Skip to main content

Full text of "Spatial accuracy assessment in natural resources and environmental sciences : second international symposium"

See other formats


\ 


Historic,  Archive  Document 

Do  not  assume  content  reflects  current 
scientific  knowledge,  policies,  or  practices. 

■ 


S91) 


.  United  States  P*-****^ 
)  Department  of 
Agriculture 

Forest  Service 


Rocky  Mountain 
Forest  and  Range 
Experiment  Station 


Spatial  Accuracy  Assessment  in 

Natural  Resources  and 
Environmental  Sciences: 

Second  International  Symposium 


Fort  Collins, 
Colorado  80526 


General  Technical 
Report  RM-GTR-277 


May  21-23, 1996 
Fort  Collins,  Colorado 


H.  Todd  Mowrer,  Raymond  L.  Czaplewski,     H.  Hamre 


Abstract 


Mowrer,  H.  Todd;  Czaplewski,  Raymond  L.;  Hamre,  R.  H.  tech  coords.  1996. 
Spatial  Accuracy  Assessment  in  Natural  Resources  and  Environmental  Sciences: 
Second  International  Symposium.  May  21-23,  1996.  General  Technical  Report 
RM-GTR-277.  Fort  Collins,  CO:  U.S.  Department  of  Agriculture,  Forest  Service, 
Rocky  Mountain  Forest  and  Range  Experiment  Station.  728  p. 

This  international  symposium  on  theory  and  techniques  for  assessing  the 
accuracy  of  spatial  data  and  spatial  analyses  included  more  than  ninety  presen- 
tations by  representatives  from  government,  academic,  and  private  institutions 
in  over  twenty  countries  throughout  the  world.  To  encourage  interactions  across 
disciplines,  presentations  in  the  general  subject  areas  of  spatial  statistics,  geo- 
graphic information  systems,  remote  sensing,  and  multidisciplinary  approaches 
were  intermixed  throughout  the  three  days  of  sessions. 


Primary  Sponsor 

Rocky  Mountain  Forest  and  Range  Experiment  Station 
United  States  Department  of  Agriculture,  Forest  Service 
Fort  Collins,  Colorado 


Endorsing  Organizations 

International  Union  of  Forest  Research  Organizations  (IUFRO): 
Sections  4.01,  4.02,  4.11,  and  4.12 
American  Statistical  Association:  Section  on  Statistics  and  the  Environment 
National  Center  for  Geographic  Information  Analysis 
International  Biometric  Society 
American  Society  for  Photogrammetry  and  Remote  Sensing:  GIS  Division, 
European  Association  of  Remote  Sensing  Laboratories: 
Special  Interest  Group  on  Landuse  Planning  and  Forestry 
United  States  Environmental  Protection  Agency 


Copies  of  this  publication  may  be  obtained  by  the  following  methods: 

E-mail:  /s=r.schneider/oul=s28a@mhs-fswa.attmail.com 

Phone:  (970)498-1719 

FAX:         (970)  498-1660      Attn:  Publications 

Mail:  Publications  Distribution 

Rocky  Mountain  Forest  and  Range  Experiment  Station 
3825  East  Mulberry  Street 
Fort  Collins,  CO  80524-8597 


USDA  Forest  Service 

General  Technical  Report  RM-GTR-277 


Spatial  Accuracy  Assessment  in 
Natural  Resources  and  Environmental  Sciences: 

Second  International  Symposium 

Technical  Coordinators: 
H.  Todd  Mowrer,  Raymond  L.  Czaplewski,  R.  H.  Hamre 
Rocky  Mountain  Forest  and  Range  Experiment  Station1 


1  Headquarters  is  in  Fort  Collins,  in  cooperation  with  Colorado  State  University. 


Preface 


This  Second  International  Symposium  on  Spatial  Accuracy  Assessment  in 
Natural  Resources  and  Environmental  Sciences  follows  two  years  after  the 
first  symposium  in  Williamsburg,  Virginia,  USA,  chaired  by  Dr.  James  L. 
Smith.  We  hope  to  maintain  and  build  on  the  spirit  of  open  discussion  and 
considerate  but  honest  evaluation  that  Jim  engendered  throughout  the 
Williamsburg  Symposium. 

It  is  apparent  from  the  size  of  these  proceedings  that  spatial  accuracy 
assessment  is  an  area  of  increasing  effort  and  importance.  Within  the  past 
decade,  we  have  developed  extensive  spatial  databases  and  have  begun  to 
realize  their  potential.  We  have  now  reached  a  point  of  maturity  in  spatial 
analyses  where  we  must  routinely  evaluate  the  quality  of  those  data  sets,  and 
the  consequent  reliability  of  the  end-products  of  their  use.  Spatial  data,  and 
model  predictions  of  spatial  attributes,  are  becoming  widely  recognized  as 
necessary  support  for  intelligent  decisions  in  resource  management  and 
environmental  protection.  Decision-makers  and  their  staffs  are  becoming 
more  accountable  for  evaluation  of  the  quality  of  spatial  data  and  model 
predictions,  upon  which  policy  analyses,  strategic  plans,  and  risk  assessments 
are  built.  Objective  evaluations  of  spatial  accuracy  are  necessary  to  compare 
alternative  technologies  and  competing  models.  Such  comparisons  speed 
improvement  in  our  knowledge  and  in  the  state-of-the-art. 

In  order  to  further  the  informed  use  of  spatial  data  and  to  improve  data 
quality,  the  natural  resource  and  environmental  science  communities  must  use 
a  multidisciplinary  approach,  drawing  from  the  methodologies  presented  here. 
We  have  attempted  to  build  upon  the  theme  of  the  first  symposium, 
"unlocking  the  puzzle,"  through  the  logo  of  puzzle  pieces  used  on  the  cover  of 
these  proceedings,  indicating  the  many  techniques  and  subject  areas  that  must 
be  intermeshed  to  produce  a  coherent  picture  of  spatial  accuracy  assessment. 

The  Steering  Committee  wishes  to  express  our  deep  appreciation  to  Robert  H. 
Hamre,  retired  Station  Editor,  who  kindly  volunteered  his  services  to 
professionally  format  these  proceedings.  Bob  has  generously  and  cheerfully 
given  the  proceedings  a  level  of  quality  that  otherwise  would  not  have  been 
within  our  resources.  The  papers  included  in  these  proceedings  were 
produced  in  camera-ready  format  by  the  individual  authors,  who  have  sole 
responsibility  for  their  content.  By  virtue  of  their  effort,  participation,  and 
support,  it  is  our  hope  that  these  proceedings  may  provide  a  significant 
contribution  to  further  the  recognition  and  development  of  spatial  accuracy 
assessment  in  natural  resources  and  environmental  sciences. 


Contents 


Page 

Spatial  Statistics 


Statistical  modeling  of  environmental  data  in  space  and  time   1 

Noel  Cressie 

A  case  study  in  geostatistical  modeling  for  petroleum  reservoir  description   4 

Jeffrey  M.  Yarns,  Jeffrey  A.  May,  and  Timothy  C.  Coburn 
An  overview  of  stochastic  spatial  simulation   13 

R.  Mohan  Srivastava 

Choosing  and  using  simulation  algorithms   23 

Donald  E.  Myers 

The  components  of  geostatistical  simulation   30 

Carol  A.  Gotway  and  Brian  M.  Rutherford 
A  general  method  for  the  efficient  selection  of  sampling  locations  for  problems 

in  environmental  restoration   39 

Brian  M.  Rutherford 

Consideration  of  spatial  variability  in  the  management  of  non-point  source 

pollution  to  groundwater   49 

W.  Woldt,  F.  Goderya,  M.  Dahab,  and  I.  Bogardi 
Incorporating  soil  variability  into  a  spatially  distributed  model  of  percolate 

accounting   57 

Andrew  S.  Rogowski 

Response  of  a  watershed  model  to  varying  spatial  landscape  chactesristics   65 

Dennis  P.  Swaney,  Wen  Ling  Kuo,  David  A.  Weinstein,  Charles  Mohler, 
Stephen  DeGloria,  Chris  Pelkie,  Fuan  Tsai,  Tammo  S.  Steenhuis,  and 
Charles  E.  McCulloch 

Understanding  the  spatial  distribution  of  tree  species  in  Pennsylvania   73 

Rachel  Riemann  Hershey 
A  calibration-based  model  for  correcting  area  estimates  from  coarse  resolution 

land  cover  data   83 

Aaron  Moody 

Geostatistical  analysis  of  multi-spatial  resolution  imaging  spectrometer  data  for 

characterizing  forest  ecosystems..   91 

Paul  M.  Treitz  and  Philip  J.  Howarth 
Application  of  multiscale  modeling  for  characterizing  variability  in  hydrologic 
processes   99 

Praveen  Kumar 

Flow  based  scale-up  of  heterogeneous  porous  media  using  homogenization  and 

wavelet  representation   107 

Joe  Koebbe  and  Ryan  Thomas 
Nonhomogeneous  hidden  Markov  models  allowing  stochastic  downscaling  of 

synoptic  atmospheric  patterns  to  local  hydrologic  phenomena   115 

Peter  Guttorp  and  James  P.  Hughes 
Explicit  consideration  of  multiple  landscape  scales  while  selecting 

spatial  resolutions   121 

John  B.  Collins  and  Curtis  E.  Woodcock 

(Continued) 


t  •  • 

in 


Using  trend  surface  methodology  to  compare  spatial  surfaces   129 

Dale  Zimmerman,  Zhi-Jun  Liu,  and  George  Hallberg 
The  effect  of  spatial  covariance  heterogeneity  on  prediction  variance   137 

/.  Andrew  Royle  and  Doug  Nychka 
Further  explorations  of  relationships  between  semi-variogram  and 

spatial  autoregressive  models   147 

Daniel  A.  Griffith,  Larry  J.  Lane,  and  Philip  G.  Doyle 
A  gamma-function  model  for  d-variate  spectra  and  cross-spectra  for  large  scale 

frequency  domain  simulation  of  stationary  random  functions  in  Rn   155 

Leon  F.  Borgman  and  John  W.  Kern 
Computing  the  area  affected  by  phosphorus  runoff  in  an  Everglades  wetland 

using  Bayesian  and  universal  kriging   164 

Song  S.  Qian 

Mapping  synecological  coordinates:  a  spatial  analysis  of  environmental 

indices  in  a  forested  landscape  ,   173 

Margaret  R.  Holdaway  and  Gary  J.  Brand 
On  the  robustness  of  the  data  assimilation  methods  in  air  pollution  modeling   181 

X.  F.  Zhang  and  A.  W.  Heemink 

Geographic  Information  Systems 

The  development  of  a  GIS  information  quality  module   189 

Jane  Drummond,  Allan  Brown,  Du  Daosheng,  Gome  van  Elzakker 
Communication  of  uncertainty  in  spatial  data  to  policy  makers   199 

Morwenna  Spear,  Jane  Hall,  and  Richard  Wadsworth 
Measuring  the  performance  of  algorithms  for  generating  ground  slope   208 

William  H.  Ryder  and  Demetra  E.  Voyadgis 
Experimental  development  of  a  model  of  vector  data  uncertainty   217 

G.  /.  Hunter,  J. Hock,  M.  Robey,  and  M.  F.  Goodchild 
A  user-friendly  tool  for  error  modelling  and  error  propagation  in  a 

GIS  environment   225 

Frank  Forier  and  Frank  Canters 
Discrete  polygons  or  a  continuous  surface:  which  is  the  appropriate  way  to 

model  forests  cartographically?   235 

Kim  Lowell 

Choosing  between  abrupt  and  gradual  spatial  variation?   243 

Gerard  B.  M.  Heuvelink  and  Johan  A.  Huisman 
Stratified  kriging  using  vague  transition  zones   251 

G.  Boucneau,  M.  Van  Meirvenne,  O.  Thas,  and  G.  Hofman 
The  effect  of  uncertain  locations  on  disease  cluster  statistics   259 

Geoffrey  M.  Jacquez  and  Lance  A.  Waller 
Reliability  of  area  mapping  by  delineation  in  aerial  photographs   267 

Claus-Peter  Gross  and  Petra  Adler 
The  effect  of  database  generalization  on  the  accuracy  of  the  viewshed   272 

Peter  Fisher 

Modeling  slope  stability  uncertainty:  a  case  study  at  the  H.  J.  Andrews 

Long-Term  Ecological  Research  Site,  Oregon   281 

Michelle  L.  Murillo  and  Gary  J.  Hunter 


(Continued) 


iv 


An  accuracy  assessment  of  a  spatial  bioclimatic  model   291 

D.  W.  McKenney,  B.  G.  Mackey,  M.  F.  Hutchinson,  and  R.  A.  Sims 
An  assessment  of  the  horizontal  accuracy  of  interim  terrain  data   301 

L.  A.  Fatale,  J.  A.  Messmore,  and  J.  R.  Ackeret 
Acquisition  of  spatial  data  by  forest  management  agencies:  a  review   309 

Michael  J.  C.  Weir 

Attribute  and  positional  accuracy  assessment  of  the  Murray  Darling  Basin 

project,  Australia   317 

R.  W.  Fitzgerald,  K.  T.  Ritman,  and  A.  Lewis 
A  new  method  for  evaluating  positional  map  accuracy   327 

Michael  Lodin  and  David  Skea 
Exploring  spatial  confidence  with  a  raster-based  GIS   336 

Matthew  H.  Pelkki 

GPS-located  accuracy  assessment  plots  on  the  Modoc  National  Forest   343 

Kevin  Casey 

A  method  for  measuring  the  spatial  accuracy  of  coordinates  collected  using 

the  global  positioning  system   351 

Thomas  Owens  and  David  McConville 
GPS  vs.  traditional  methods  of  data  imput:  improving  spatial  data  accuracy?   359 

Russell  Combs,  Jr.,  James  L.  Smith,  and  Paul  V.  Bolstad 

Remote  Sensing 

Large  scale  tropical  forest  change  monitoring  using  multiple  resolution 

satellite  data:  from  hot  spot  detection  to  global  deforestation  assessment?   366 

Herve  Jeanjean,  Frederic  Achard,  and  Jean-Paul  Malingreau 
Optimum  area  sampling  frame  using  high  resolution  satallite  images  with 
operational  objective:  how  to  conciliate  statistical  requirements  and 

practical  aspects   375 

Helene  De  Boissezon  and  Herve  Jeanjean 

The  use  of  AVHRR  satellite  imagery  to  monitor  boreal  ecosystem  forest  fires   383 

Donald  R.  Cahoon,  Jr.,  and  Brian  J.  Stocks 
Large  area  forest  cover  assessment:  effects  of  misregistration  in  a  double 

sampling  approach  with  coarse  and  high  resolution  satellite  images   391 

Christoph  Kleinn,  Berthold  Traub,  and  Matthias  Dees 

Linear  mixture  modeling  with  autocorrelated  errors   401 

Jayantha  Ediriwickrema,  Siamak  Khorram,  Marcia  Gumpertz,  and 
John  Brockhaus 

Validation  process  of  the  USGS  seasonal  land  cover  regions  of  the 

conterminous  United  States   409 

Zhiliang  Zhu,  Donald  O.  Ohlen,  Raymond  L.  Czaplewski,  and 
Robert  E.  Burgan 

Investigation  of  possible  contributions  NDVTs  have  to  misclassification  in 

AVHRR  data  analysis   419 

David  L.  Evans  and  Raymond  L.  Czaplewski 
Measures  of  association  and  agreement  for  describing  land  cover 

characterization  classes   425 

Eugene  A.  Fosnight  and  Gary  W.  Fowler 

Forest  spatial  surveys  using  the  Rao-Hartley-Cochran  sampling  design   434 

Jeffrey  S.  Pontius 

(Continued) 


v 


Increasing  spatial  precision  and  accuracy  for  monitoring  peatlands  in 

Switzerland  by  remote  sensing  techniques   441 

Michael  Kohl,  Hans  Jbrg  Schnellbacher,  and  Andreas  Griinig 
Spatial  and  temporal  models  in  contextual  classification   451 

Bo  Ranneby 

Generalized  linear  mixed  models  for  analyzing  error  in  a  satellite-based 

vegetation  map  of  Utah   459 

Gretchen  G.  Moisen,  D.  Richard  Cutler,  and  Thomas  C.  Edwards,  Jr. 
Statistical  properties  of  measures  of  association  and  the  kappa  statistic  for 

assessing  the  acccuracy  of  remotely  sensed  data  using  double  sampling   467 

Mohammed  A.  Kalkhan,  Robin  M.  Reich,  and  Raymond  L.  Czaplewski 
Comparing  methods  used  to  combine  multi-level  samples  of  compositional  data ....  477 

Murray  Todd  Williams 
Cost-effective,  practical  sampling  strategies  for  accuracy  assessment  of 

large-area  thematic  maps   485 

Stephen  V.  Stehman 

Spatial  and  probabilistic  classification  of  forest  structures  using  landsat  TM  data...  493 

Jeffrey  L.  Moffett  and  Julian  Besag 
Total  error  estimation  in  a  spatial  database  for  GIS   501 

Jose  Alberto  Quintanilha  and  Marcos  Rodrigues 
Assessing  the  accuracy  of  a  regional  land  cover  classification   508 

William  Gierke,  Raymond  Czaplewski,  Jeff  Campbell,  and  Janet  Fahringer 
Sampling  satellite  images  for  area  estimates  in  a  large  region   509 

F.  /.  Gallego 

Cooperative  accuracy  assessment  strategies  for  sampling  a  natural  landcover 

map  of  Arkansas   517 

Robert  S.  Dzur,  Michael  E.  Garner,  Kimberly  G.  Smith,  W.  Fredrick  Limp, 
Donald  G.  Catanzaro,  and  Richard  L.  Thompson 

Multidisciplinary 

Fuzzy  measures  in  multi-criteria  evaluation   527 

/.  Ronald  Eastman  and  Hong  Jiang 
On  roles  and  goals  for  map  accuracy  assessment:  a  remote  sensing  perspective   535 

Curtis  E.  Woodcock 
Integration  of  inventory  and  field  data  for  automated  fuzzy  accuracy 
assessment  of  large  scale  remote-sensing  derived  vegetation  maps  in 
Region  5  national  forests   541 

Jeff  A.  Milliken  and  Curtis  E.  Woodcock 
Uncertainty  of  spatial  metric  relations  in  GIS   545 

Xiaoyong  Chen,  Takeshi  Doihara,  and  Mitsuru  Nasu 
Plot  collocation  error:  impacts  on  area  estimation   553 

Willem  W.  S.  van  Hees 
Sensitivity  of  landscape  pattern  metrics  to  misclassification  and  differences 

in  land  cover  composition   560 

James  D.  Wickham,  Robert  V.  O'Neill,  Kurt  H.  Riitters,  Timothy  G.  Wade, 

and  K.  Bruce  Jones 

Spatial  (in)consistency  of  watershed  delineations  among  agencies  and  scales  in 

Pennsylvania   561 

Wayne  L.  Myers,  Barry  M.  Evans,  and  Michael  C.  Anderson 

(Continued) 

vi 


Communicating  the  results  of  accuracy  assessment:  Metadata,  ditital  libraries, 

and  assessing  fitness  for  use   568 

Michael  F.  Good  child 

Covariate-directed  sampling  for  assessing  species  richness   569 

G.  P.  Patil,  Glen  Johnson,  and  Matteo  Grigoletto 
Data  accuracy  to  data  quality:  using  spatial  statistics  to  predict  the  implications 

of  spatial  error  in  point  data   577 

Adam  Lewis  and  Michael  F.  Hutchinson 
Moving  into  secondary  map  projections:  an  analysis  of  potential  inconsistencies 

in  spatial  data   585 

Mohamad  N.  Said  and  Peter  F.  Fisher 
Filling  in  missing  forestry  data:  exploring  autocorrelational  techniques   593 

Alissa  N.  Antle  and  Peter  L.  Marshall 
Optimizing  sampling  schemes  for  mappping  and  dredging  polluted 

sediment  layers   601 

L.  Hazelhoff  and  F.  Hoefsloot 
Forest  cover  monitoring  in  India:  the  satellite  experience   610 

/.  B.  Lai 

Study  about  the  use  of  SPOT  data  for  estimation  of  pastoral  areas  in  Morocco   617 

Philippe  Brion 

The  influence  of  vegetation  cover  density  and  topographic  parameters  on  the 

thermal  emission  of  the  beech  forests  of  Simbruini  Mountains  (central  Italy)   623 

Carlo  Ricotta,  Giancarlo  Avena,  and  Fernando  Ferri 
Application  of  non-parametric  kernel  regression  and  nearest-neighbor  regression 

for  generalizing  sample  tree  information  «   631 

Annika  Kangas  and  Kari  T.  Korhonen 
A  method  for  assessing  the  prediction  quality  of  mechanistic  forest 

growth  models   639 

Biing  T.  Guan,  George  Gertner,  and  Pablo  Parysow 
Statistical  analysis  of  error  propagation  in  national  level  carbon  budgets   649 

C.  /.  Cieszewski,  D.  P.  Turner,  and  D.  L.  Phillips 
Monte  Carlo  simulations  of  nonlinear  size-age  relationships   659 

Ronald  E.  McRoberts 

Cross-correlations  among  single  tree  growth  models   667 

Hubert  Hasenauer,  Robert  A.  Monserud,  and  Timothy  G.  Gregoire 
Genetics,  geographies,  and  prairie  dogs:  error  and  accuracy  in  a  validated 

spatially-explicit  dispersal  model   676 

Gillian  Bowser 

Effect  of  uncertainty  in  mapped  biodiversity  data  on  optimal  conservation 

decisions   685 

Michael  J.  Conroy  and  Jennifer  E.  Crocker 
Sensitivity  analysis  of  species  richness  mapping  to  variations  in  forest-wildlife 

relationships   693 

Tom  Kohley 

Development  of  a  survey  sampling  methodolgy  for  rare  species   694 

Molly  Van  Caster,  David  Bowden,  and  Jennifer  Hoeting 
Methods  to  analyse  the  spatial  structure  of  plant  communities   695 

Paul  Braun,  Heiko  Balzter,  and  Wolfgang  Kohler 
Modeling  population  dynamics  with  cellular  automata   703 

Heiko  Baltzer,  Paul  Braun,  and  Wolfgang  Kohler 

(Continued) 


vii 


Posters 


Inconsistency  of  line  convolution  as  a  quality  factor  in  them  aintenance  of 

spatial  databases   713 

Michael  J.  C.  Weir 

Lead-based  paint  survey  data:  spatial  variation   714 

Tze-San  Lee 

Uncertainty  of  spatially-averaged  rainfall  estimates  from  rain  gauges   715 

Jeffrey  R.  McCollum  and  Witold  F.  Krajewski 
Defining  urban  settlements:  certainties  regarding  uncertainties   716 

Mahavir  and  Sahar  Al-Amir 
Spatial  dependence  of  tree  biomass  in  an  old-growth  forest   717 

Franco  Biondi  and  Donald  E.  Myers 
Sensibility  and  uncertainty  analyses  of  an  expert  system  to  determine 

stand  treatments   718 

Otto  Eckmuellner  and  Martin  Moser 
Stratified  two-stage  sampling  (self-weighted)  for  assessment  of  village 

forest  resources  in  Bangladesh   719 

S.  S.  Islam 

Modelling  animal  movements  within  a  GIS  framework   720 

/.  Joachim,  G.  Janeau,  and  F.  Spitz 
Evaluation  of  forest  condition  assessment  data   721 

F.  Mitterbock  and  O.  Eckmiillner 
Application  of  remote  sensing  technology  in  assessing  village  wood 

resources  in  Bangladesh   722 

Jamil  Ahmed  Chowdhury 
Quality  evaluation  services  on  the  Internet   723 

Anders  Ostman 

Assessing  attainment  of  standard  by  two  stage  sampling   724 

Debashis  Kushary 

Site  positioning  accuracies  using  GPS:  Y-code  receivers  and  real-time 

differential  options   725 

Karl  E.  Brown 

Statistical  models  of  landscape  pattern  and  the  effects  of  coarse  resolution  of 

satellite  imagery  on  estimation  of  area   726 

Christine  A.  Hlavka 

Author  index   727 


viii 


Statistical  Modeling  of 
Environmental  Data  in  Space  and 

Time 

Noel  Cressie1 


EXTENDED  ABSTRACT 

As  a  concept,  the  environment  is  simply  the  surroundings  of  an  organism 
or  organisms.  Space  and  time  scales  in  environmental  investigations  can  range 
from  the  very  local  to  the  very  global.  Some  studies  attempt  to  understand 
physical,  chemical,  and  biological  processes  by  performing  controlled  experi- 
ments in  the  laboratory.  In  this  paper,  we  are  concerned  instead  with  studies 
made  in  the  field.  These  are  mostly  observational  in  nature  and,  hence,  even 
though  a  large  amount  of  data  may  be  collected  and  analyzed,  one  typically  can 
only  infer  associations  rather  than  causation. 

Most  environmental  studies  in  the  field  involve  variability  over  both  space 
and  time.  The  extension  of  traditional  geostatistical  methods,  such  as  kriging, 
to  the  space-time  domain  is  one  possible  approach  to  characterizing  the 
variability  of  the  processes  (e.g.,  Bilonick,  1983;  Eynon  and  Switzer,  1983; 
Stein,  1986;  Le  and  Petkau,  1988;  Sampson  and  Guttorp,  1992;  Host  et  al., 
1995).  There  are  difficult  modeling  decisions  to  make  in  this  approach, 
involving  space,  time,  and  space-time  interaction  components.  The  space-time 
variability  is  characterized  by  a  variogram  that  often  exhibits  very  different 
spatial  behavior  at  different  points  in  time  and  the  class  of  variogram  models 
that  can  be  fit  in  this  situation  is  very  small  indeed.  In  the  atmospheric 
sciences,  traditional  methods  for  examining  space-time  processes  have  focused 
on  Empirical  Orthogonal  Functions  (EOF),  Canonical  Correlation  Analysis 
(CCA),  and  Principal  Oscillation  Patterns  (POP);  see,  for  example,  von  Storch 
et  al.  (1995).  Although  these  techniques  are  visually  powerful,  they  were 
designed  with  summarization  rather  than  prediction  in  mind. 

Without  the  spatial  component,  there  are  a  large  class  of  time  series  that 
could  be  used  to  model  the  temporal  component  (e.g.,  autoregressive  error 
processes).  These  are  dynamic  in  the  sense  that  they  exploit  the  unidirectional 
flow  of  time.  Without  the  temporal  component,  geostatistical  methods  could  be 
used  to  model  the  spatial  component  (e.g.,  intrinsically  stationary  error 
processes).  These  are  descriptive  in  the  sense  that  although  they  model  spatial 
correlation  there  is  no  causative  interpretation  associated  with  them.  When  both 


Noel  Cressie  is  Professor  of  Statistics  and  Distinguished  Professor  in  Liberal  Arts  and  Sciences,  Iowa  State  University, 
Ames,  I  A  50011. 


1 


temporal  and  spatial  components  are  present,  it  seems  sensible  to  use  models 
that  are  a  combination  of  both  approaches,  namely  temporally  dynamic  and 
spatially  descriptive.  That  is  the  new  feature  of  our  work  and  it  allows  a  natural 
development  of  the  space-time  Kalman  filter. 

To  give  some  definiteness  to  the  problem,  consider  Markov  temporal 
models  with  spatial  colored  noise: 

5t(s)  =  0^(5)  +  a25t.2(s)  +  ...  +  apSt.p(s)  +  ^(s);  selR2,  (1) 

where  St(s)  is  an  (unobserved)  value  of  the  state  process  at  location  s  and  time 
t.  The  observations  are  actually 

Us)  =  5t(s)  +  et(s),  (2) 

which  expresses  the  data  as  a  noisy  version  of  the  state  process.  The  goal  is  to 
predict  St()(so)>  where  both  to  and  s0  may  or  may  not  represent  space-time 

coordinates  at  which  data  are  available.  In  (1),  (<%,..., ap)  are  autoregressive 
parameters  (i.e.,  parameters  of  the  "temporally  dynamic"  component)  and,  most 
importantly,  r]t(-)  is  a  spatially-colored  noise  process  (i.e.,  the  "spatially 
descriptive"  component).  In  (2),  et(-)  is  a  white-noise  process  representing 
measurement  error. 

Data  come  in  the  form, 

Zj  -  (Zfa),...,^)';]  =  l,...,t, 

where  it  is  not  essential  that  all  observations  are  available  at  each  time  point  and 
at  each  spatial  location.  The  optimal  predictor  of  £t  (Sq)  is: 

£0|t(So)  -  EP^IZ^ZJ, 
with  mean-squared  prediction  error, 

E(5to|t(s0)-5to(So))2. 

Both  these  quantities  can  be  calculated  recursively,  using  what  could  be  called 
a  space-time  Kalman  filter  (Huang  and  Cressie,  1996). 

The  state-space  model  in  (1)  is  very  attractive  because  it  features  the 
dynamic  aspect  through  an  autoregressive  structure  but  builds  in  space-time 
interaction  through  the  error  process  rjt,  which  is,  at  any  point  in  time,  a 
spatially  correlated  (e.g.,  intrinsically  stationary)  process.  Notice  that  5t(s)  is 
influenced  directly  by  past  values  only  at  location  s.  In  reality,  spatio-temporal 
processes  are  likely  to  be  more  complicated,  to  have  dependence  also  on  past 
values  at  locations  u  near  s.  Thus,  I  shall  investigate  the  spatio-temporal 
climate  model, 


2 


S,(s)  =  a  J  a>,(u)S,_,(u)dii  +  tj,(s), 


(3) 


where,  for  identifiability,  the  coefficients  cos(u)  satisfy  J  ws(u)du  =  1. 
Development  of  the  spatio-temporal  Kalman  Filter  allows  prediction  of  Sl  (s0) 
based  on  data  ZX,...,ZV 

The  two  challenging  aspects  of  this  research  are  to  derive  the  expressions 
for  the  optimal  predictor  and  the  optimal  mean-squared  prediction  error,  and  to 
obtain  efficient  estimators  of  the  parameters  a,  var(et(s)),  and  cov(T7t(s),?7t(u)). 
These  are  then  substituted  into  the  optimal  prediction  equations.  (This  research 
is  in  progress  with  Ph.D.  student  Christopher  K.  Wikle.)  An  alternative  to  the 
estimation  of  parameters  is  to  put  (prior)  distributions  on  them.  In  this  case, 
there  is  a  good  physical  reason  to  do  this.  The  parameters  almost  certainly  vary 
from  year  to  year  and  from  region  to  region;  this  extra  variation  can  quite 
simply  be  handled  by  replacing  a  in  (3)  with  {a0,...,a^,at,...}  and  assuming 
them  to  be  distributed  according  to  some  prior  distribution.  The  goal  is  still  to 
obtain  E(StQ(sQ)\Zl,...,Z^)  and  its  mean-squared  prediction  error  or,  more 

generally,  the  posterior  distribution  of  5t  (s0)  given  Zu...,Zt.    While  such 

calculations  were  daunting  five  years  ago,  Markov  chain  Monte  Carlo  (e.g., 
Bernardo  and  Smith,  1994,  Section  5.5.5)  can  be  invoked  to  handle  the 
problem. 

REFERENCES 

Bernardo,  J.M.  and  Smith,  A.F.M.  (1994).  Bayesian  Theory.  Wiley, 
Chichester. 

Bilonick,  R.  A.  (1983).  Risk  qualified  maps  of  hydrogen  ion  concentration  for 
the  New  York  state  area  for  1966-1978.  Atmospheric  Environment,  17, 
2513-2524. 

Eynon,  B.P.  and  Switzer,  P.  (1983).    The  variabiity  of  rainfall  acidity. 

Canadian  Journal  of  Statistics,  11,  11-24. 
Host,  G.,  Omre,  H.,  and  Switzer,  P.  (1995).  Spatial  interpolation  errors  for 

monitoring  data.  Journal  of  the  American  Statistical  Association,  90,  853- 

861. 

Huang,  H.C.  and  Cressie,  N.  (1996).  Spatio-temporal  prediction  of  snow  water 
equivalent  using  the  Kalman  filter.  Computational  Statistics  and  Data 
Analysis,  in  press. 

Le,  D.N.  and  Petkau,  A.J.  (1988).  The  variability  of  rainfall  acidity  revisited. 

Canadian  Journal  of  Statistics,  16,  15-38. 
Sampson,  P.D.  and  Guttorp,  P.  (1992).     Nonparametric  estimation  of 

nonstationary  spatial  covariance  structure.     Journal  of  the  American 

Statistical  Association,  87,  108-119. 
Stein,  M.L.  (1986).   A  simple  model  for  spatial  temporal  processes.  Water 

Resources  Research,  22,  2107-2110. 
von  Storch,  H.,  Burger,  G.,  Schnur,  R.,  and  von  Storch,  J.S.  (1995).  Principal 

oscillation  patterns:  A  review.  Journal  of  Climate,  8,  377-400. 


3 


A  Case  Study  in  Geostatistical 
Modeling  for  Petroleum  Reservoir 

Description 

Jeffrey  M.  Yarus1,  Jeffrey  A.  May2,  and  Timothy  C.  Coburn3 


Abstract.—Geostatistical  techniques  can  be  used  to  effectively 
characterize  the  subsurface  properties  of  petroleum  reservoirs,  even  in 
the  face  of  limited  amounts  of  control  data.  The  applications 
described  here  pertain  to  a  newly-discovered  offshore  oil  field  in 
southeast  Asia,  the  extent  of  which  covers  approximately  35  square 
kilometers.  The  study  involves  the  development  of  a  three- 
dimensional  grid  of  estimated  rock  properties  (e.g.,  porosity  and 
lithology),  consisting  of  approximately  3.1  million  nodes,  to  be  used 
in  computer  simulations  of  field  production  performance.  A  project 
of  this  scope  requires  a  novel  combination  of  modeling  procedures  and 
spatial  analysis  tools,  including  embedded  Markov  chain  analysis, 
sequential  Gaussian  simulation,  and  probability  field  simulation.  A 
description  of  the  overall  approach  is  presented,  along  with  an 
example  of  the  modeling  results.  


INTRODUCTION 

Petroleum  engineers  use  computer  simulations  to  establish  the  likely  production 
profiles  of  oil  and  gas  fields  and  reservoirs.  Such  simulations  are  based  on  a 
number  of  input  parameters  which  are  estimated  through  empirical  observation, 
direct  and  indirect  measurement  in  the  subsurface,  or  by  analogy  to  similar 
hydrocarbon  bearing  provinces.  The  resulting  profiles  are  used  to  guide  further 
development  of  fields  through  drilling  of  additional  wells  or  by  the  adoption  of 
enhanced  extraction  or  recovery  methods. 

The  data  needed  to  estimate  the  input  parameters  for  the  computer  simulations 
includes  both  qualitative  and  quantitative  descriptions  of  the  geological  regime, 
downhole  measurements  (acoustic,  electrical,  etc.)  of  the  subsurface  media, 
laboratory  measurements  on  soil,  rock,  and  fluid  samples,  seismic  reflections, 
drilling  records,  and  subjective  geological  experience  and  intuition.  All  too  often, 

1  Consulting  Mathematical  Geologist,  Yarus  &  Associates,  Denver,  CO. 

2  Geological  and  Geophysical  Consultant,  Littleton,  CO. 

3  Senior  Statistician,  National  Renewable  Energy  Laboratory,  Golden,  CO. 


4 


however,  many  of  the  requisite  data  are  missing  or  simply  unavailable,  a  situation 
which  is  particularly  true  in  the  case  of  newly-discovered  fields,  or  in  foreign 
and/or  offshore  settings,  where  data  collection  can  be  a  comparatively  expensive 
and  logistically  difficult  proposition. 

Historically,  the  two  most  common  ways  to  remedy  missing  data  have  been  to 
impute  it  from  knowledge  of  other  producing  fields  in  geologically  similar 
environments  (best  case)  or  to  simply  "make  it  up"  (worst  case).  As  an 
alternative,  the  discipline  and  tools  of  geostatistics  have  recently  been  shown  to 
provide  a  more  rigorous  and  statistically-reliable  means  of  estimating  the 
simulation  parameters  by  taking  advantage  of  the  notion  of  spatial  correlation  that 
is  so  central  to  the  investigation  of  all  earth  processes  and  phenomena. 

This  paper  discusses  the  effectiveness  of  the  geostatistical  approach  in  a  laterally 
extensive  geographic  setting,  and  with  the  additional  limitation  of  minimal  control 
points  and  other  tangible  data.    A  number  of  spatial  analysis  tools,  including 
embedded  Markov  chain  analysis,  conditional  simulation,  and  probability  field 
simulation  are  combined  in  novel  ways  to  produce  a  three-dimensional  grid  of 
parameter  estimates  encompassing  approximately  3.1  million  nodes.  A 
description  of  the  overall  modeling  approach  is  presented,  along  with  an  example 
of  the  modeling  results. 

STUDY  SETTING 

The  present  study  involves  a  newly-discovered  offshore  oil  field  in  southeast  Asia 
for  which  a  plan  of  development  is  necessary  to  satisfy  the  contractual 
requirements  of  the  business  concession.  Three  successful  wells  (A,  B,  and  C) 
have  been  drilled.  Drilling  records,  along  with  preliminary  seismic  data  and  initial 
production  tests  on  the  wells,  confirm  the  presence  of  hydrocarbons  in  an  area 
covering  approximately  35  square  kilometers.  Because  of  the  geographical  extent 
of  the  field,  and  the  high  cost  of  drilling  additional  wells,  it  is  desirable  to  develop 
the  most  precise  model  of  the  reservoir  possible,  prior  to  selecting  additional 
drilling  sites  or  production  platform  locations.  Such  a  model  must  reliably 
characterize  the  degree  and  lateral  extent  of  all  reservoir  properties  which  affect 
fluid  flow  and  hydrocarbon  production. 

Conceptual  Geological  Framework 

Preliminary  regional  studies  of  the  field  area  identified  the  presence  of  an  ancient 
rift  system  filled  with  clastic  sediment,  sourcing  from  the  northeast.  The 


5 


hydrocarbon-bearing  interval  (the  reservoir)  of  the  field  is  located  in  one  particular 
subsurface  horizon,  which  was  apparently  deposited  early  in  the  rift  fill  sequence. 
Using  this  information,  a  conceptual  model  was  developed  which  establishes  the 
depositional  environment  as  being  a  fluvial  dominated,  fan  delta  system. 

Available  Data 

The  data  set  consists  of  ten  seismic  lines  from  an  early  2-D  survey  over  the 
region,  and  information  from  three  wells  drilled  along  a  general  northwest-to- 
southeast  transect  in  the  central  southeast  quadrant  of  the  study  area,  the  wells 
being  spaced  at  intervals  of  approximately  2000  meters.  During  the  drilling 
operation,  cores  were  cut  from  two  of  the  wells  and  sent  to  the  laboratory  to 
support  rock  studies.  Conventional  suites  of  wireline  logs  (including  sonic, 
density,  resistivity,  neutron,  and  gamma  ray)  were  run  in  all  three  wells. 


Figure  l.—Relative  locations  of  wells,  pseudo  wells,  seismic  lines,  and  faults 


Values  of  reservoir  parameters  at  the  well  locations  were  determined  from 
laboratory  measurements  and  examination  of  the  core  material,  as  well  as  from 
numerical  analysis  of  the  well  logs.  In  particular,  individual  values  of  porosity 
were  obtained  for  each  half-foot  segment  of  the  reservoir.  Similarly,  each  half-foot 
segment  was  classified  into  one  of  four  primary  lithologies,  or  rock  types— channel 
sands,  delta  front  sands,  overbank  sands  and  muds,  and  lacustrine  muds—identified 
by  the  project  geologists.  Of  these  four  lithologies,  the  channel  sands  are 
presumed  to  have  the  greatest  potential  for  production.  The  resulting  collection  of 


6 


porosity  measurements  and  lithology  classifications  comprises  approximately 
2,500  data  pairs  per  well. 

Based  on  further  studies  of  the  cores,  analysis  of  the  wireline  logging  runs,  and 
interpretation  of  the  seismic  lines,  the  following  additional  conclusions  were 
formulated:  (1)  the  field  is  bounded  by  at  least  two  major  faults,  with  Well  B 
being  situated  outside  these  boundaries  (Figure  1);  (2)  drainage  is  controlled,  in 
part,  by  east-west  trending  faults;  (3)  four  major  depositional  cycles  (Intervals  1 
to  4),  separated  by  non-permeable  shale  barriers,  can  be  identified  which 
effectively  subdivide  the  total  reservoir  into  separate  producing  units  of  varying 
thickness;  (4)  the  principle  orientation  of  the  channel  sands,  the  dominant 
reservoir  lithology,  is  also  east  to  west;  (5)  the  geometry  of  the  channel  sands  is 
represented  by  an  average  length-to-width  ratio  of  approximately  10:1  (taken  in 
part  from  results  reported  by  Fielding  and  Crane  (1987)  and  Knutson  and  Board- 
man  (1978));  and  (6)  the  median  estimated  thickness  of  the  channel  sands  is  6  ft. 

EXPANDING  AERIAL  DATA  COVERAGE  THROUGH  PSEUDO  WELL 

DEVELOPMENT 

By  themselves,  the  three  wells  do  not  constitute  an  adequate  geographic  sampling 
of  the  reservoir  from  which  to  develop  a  three-dimensional  model.  Therefore,  in 
order  to  increase  the  aerial  distribution  of  "sampling"  locations  and  to  facilitate 
geostatistical  interpolation,  fourteen  pseudo  wells  were  introduced.  Using  the 
process  described  below,  these  pseudo  wells  were  created  in  such  a  way  as  to 
reflect  the  nature  of  the  reservoir  observed  at  the  actual  drillsite  locations,  adjusted 
for  the  features  of  the  conceptual  geological  framework,  relationships  extracted 
from  geological  outcrop  studies,  and  trends  observed  in  the  seismic  data. 

Pseudo  well  sites  were  determined  to  coincide  with  several  of  the  shot  point 
locations  identified  along  the  ten  seismic  lines,  resulting  in  the  "sampling"  coverage 
illustrated  in  Figure  1 .  The  shot  point  locations  provided  a  convenient  additional 
data  source  in  that  the  seismic  velocity  recorded  at  these  sites  could  be  used  to 
calculate  acoustic  impedance  (AI),  a  continuous  reservoir  property  from  which  the 
boundaries  of  the  four  major  depositional  cycles  could  be  estimated.  To  do  so,  log 
suites  and  synthetic  seismograms  from  the  three  wells  were  manually  depth- 
shifted  to  obtain  an  optimum  visual  alignment  of  all  petrophysical  markers.  This 
identification  of  the  major  cycle  boundaries  was  corroborated  through  an 
automated  "correlation"  approach  (Olea,  1988).  The  depths  associated  with  these 
boundaries  were  then  projected  to  the  acoustic  impedance  trace  at  each  pseudo 
well  location. 


7 


a.  Set  of  Markov  sequences  modeled 
tor  one  interval 

b.  At  Trace 
in 


CONCEPTUAL  GEOLOGICAL  MODEL 


UNCONNNED 
£LOW  &  BRAIDED 
CHANNELS 


Pseudo-Well  Location 


c.  Stacked  Fan  Delta 
comprising  an  interval 


Figure  2.— The  process  for  determining  lithology  in  pseudo  wells 


Markov  Modeling  of  Lithology 


The  second  stage  of  pseudo  well  construction  involved  overprinting  a  vertical 
sequence  of  lithologies  at  each  location.  Embedded  Markov  chain  analysis  (Davis, 
1986)  was  used  to  develop  these  sequences  based  on  the  transition  probabilities 
(bottom  to  top)  observed  in  the  actual  wells.  For  illustration  purposes,  Table  1 
shows  the  transition  probabilities  for  Interval  1  computed  over  the  three  wells. 

Table  l.--Interval  1  transition  probabilities  (bottom  to  top)  for  the  four  lithologies 

To 

Lithology         Channel       Delta  Front    Overbank  Sands  Lacustrine 


Sands 

Sands 

&  Muds 

Muds 

Channel 

Sands 

0 

0 

.78 

.22 

Delta  Front 

Sands 

.28 

0 

0 

.72 

From 

Overbank 

Sands  & 

.67 

0 

0 

.33 

Muds 

Lacustrine 

Muds 

.19 

.82 

0 

0 

Applying  the  marginal  vectors,  seven  different  lithologic  sequences  were  generated 


8 


for  each  of  the  four  intervals  at  each  pseudo  well  location.  Stacking  the  interval 
sequences  together  resulted  in  28  possibilities  per  pseudo  well,  from  which  the 
most  likely  case  was  selected  by  subjective  comparison  to  the  respective  acoustic 
impedance  depth  traces,  giving  due  consideration  to  the  conceptual  geological 
model.  The  result  of  this  effort  was  the  assignment  of  a  lithology  classification  to 
each  half-foot  depth  interval  in  each  of  the  pseudo  wells  (Figure  2). 

Determination  of  Porosity  Via  Sequential  Gaussian  Simulation 

The  third  stage  of  pseudo  well  construction  involved  determination  of  a  value  of 
porosity  for  each  half- foot  depth  interval.  This  task  required  two  critical  inputs: 
(1)  the  cumulative  distribution  (CDF)  of  porosity  for  each  lithology  within  each 
of  the  four  major  intervals,  combining  information  across  wells;  and  (2)  a  vertical 
correlogram  of  porosity  for  each  of  the  four  intervals  (combining  information 
across  wells  and  ignoring  lithology).  The  CDFs  were  used  to  guarantee  that  the 
statistical  distribution  of  porosity  values  in  the  pseudo  wells  matched  that  of 
porosity  values  in  the  original  wells;  and  the  correlogram,  which  measures  the 
degree  to  which  porosity  changes  from  one  depth  to  another,  was  used  to  insure 
that  the  vertical  spatial  distributions  of  porosity  in  the  pseudo  wells  and  the 
original  wells  were  aligned. 

Given  the  CDFs  and  correlograms  described  above,  porosity  values  for  each 
individual  depth  in  each  pseudo  well  were  determined  using  sequential  Gaussian 
simulation  (Deutsch  and  Joumel,  1992),  a  form  of  conditional  simulation,  in  the 
following  way.  For  each  pseudo  well,  a  vertical  depth  was  selected  at  random 
from  its  vertical  section.  The  interval  and  lithology  associated  with  that  depth 
were  identified,  and  a  porosity  value  was  then  randomly  selected  employing  the 
appropriate  CDF.  This  Monte-Carlo-like  approach  was  successively  repeated  for 
every  depth  in  the  selected  well.  When  a  selected  depth  was  located  near  one  that 
had  been  previously  simulated  (i.e.,  within  the  range  of  correlation  defined  by  the 
appropriate  correlogram),  and  the  two  depths  were  associated  with  the  same 
lithology,  the  value  of  porosity  at  the  first  depth  influenced  determination  of  the 
value  at  the  second.  The  degree  of  influence  exerted  was  dependent  on  the 
correlation  coefficient  defmed  by  the  correlogram  for  the  distance  (in  depth) 
separating  the  two. 

3-D  PROPAGATION  OF  LITHOLOGY  AND  POROSITY  USING 
PROBABILITY  FIELD  SIMULATION 

Using  information  from  the  three  wells  and  the  fourteen,  fully-populated  pseudo 
wells,  a  full  three-dimensional  model  of  the  reservoir  was  developed.  The  initial 


9 


step  was  to  construct  a  rectangular  three-dimensional  network  grid  consisting  of 
approximately  3.1  million  nodes.  Grid  dimensions  were  set  at  180x  x  87y  x  200z 
cells,  with  individual  cell  dimensions  being  x  =  150  ft,  y  =  150  ft,  and  z  =  5  ft  (the 
z-dimension  represents  the  thickness  of  one  reservoir  layer).  A  grid  of  this 
density  was  deemed  necessary  to  depict  the  detailed  lithologic  changes 
characteristic  of  this  reservoir. 

Probability  field  simulation  (Srivastava,  1992)  was  used  to  propagate  porosity 
values  and  lithology  classifications  from  the  wells  and  pseudo  wells  to  the 
individual  grid  node  locations.  Separate  simulations  were  conducted  for  the  two 
quantities,  the  results  of  which  were  subsequently  combined  for  purposes  of 
describing  the  reservoir  in  a  multivariate  sense. 

Like  sequential  Gaussian  simulation,  the  use  of  probability  field  simulation 
(another  form  of  the  more  general  technique  of  conditional  simulation)  required 
information  about  the  statistical  distribution  of  the  reservoir  properties,  as 
characterized  by  their  CDFs,  as  well  as  spatial  variation  in  both  the  vertical  and 
lateral  directions,  as  characterized  by  the  appropriate  correlograms.  Because 
lithology  is  not  continuous,  CDFs  and  directional  correlograms  for  this  quantity 
were  determined  using  an  indicator  variable  (having  values  1  or  0)  which  designates 
the  presence  or  absence  of  the  specific  lithology  in  question. 

In  the  case  of  porosity,  spatial  variation  was  represented  in  the  vertical  sense 
using  the  same  correlogram  (described  above)  used  to  establish  porosity  values  in 
the  pseudo  wells.  To  characterize  the  spatial  variation  in  the  lateral  sense, 
additional  directional  correlograms  for  each  major  interval  were  constructed  using 
the  porosity  values  available  in  all  three  wells  and  fourteen  pseudo  wells. 
Similarly,  both  vertical  and  lateral  correlograms  were  developed  for  each  lithology 
within  each  major  interval  (combining  information  across  all  wells  and  pseudo 
wells)  based  on  the  assigned  indicator  values. 

The  application  of  probability  field  simulation,  the  algorithm  for  which  is 
described  more  fully  by  Srivastava  (1992),  required  the  development  of  a  three- 
dimensional  grid  of  randomly-assigned  probabilities  (a  probability  template)  based 
on  both  the  vertical  and  lateral  correlograms  of  the  quantity  of  interest.  This  grid 
had  the  same  physical  dimensions  as  the  simulation  grid,  with  each  node  of  the 
probability  template  corresponding  to  a  node  on  that  grid.  The  probability  values 
assigned  to  the  nodes  were  then  used  to  select  values  of  porosity  or  lith-  ology 
from  the  appropriate  CDFs,  resulting  in  a  fully-populated  simulation  grid. 


10 


EXAMPLE  OF  THE  SIMULATION  RESULTS 


Based  on  the  dimensions  of  the  simulation  grid,  the  reservoir  was  arbitrarily 
subdivided  into  200  five-foot  layers  to  optimize  precision  and  to  better  take 
advantage  of  the  available  computing  resources.  All  results  were  generated  on 
Unix-based  workstations,  using  a  combination  of  proprietary  and  commercially- 
available  code. 

Individual  simulations  of  both  porosity  and  lithology  were  developed  for  each  of 
the  200  layers,  with  several  realizations  of  each  layer  being  produced.  Figure  3 
graphically  illustrates  the  results  for  one  layer  of  porosity,  depicting  a  lateral 
gradational  character  across  the  reservoir  which  closely  mimics  input  from  the 
conceptual  geological  model,  along  with  the  well  data  and  seismic  interpretations. 


Figure  3.— Simulation  results  for  porosity  in  one  reservoir  layer 

SUMMARY 


A  number  of  geostatistical  techniques  and  other  quantitative  methods,  including 
conditional  simulation,  probability  field  simulation,  and  Markov  analysis,  were 
combined  to  fully  characterize  and  model  a  large-scale  oil  reservoir.  Drawing  on 
concomitant  information  from  a  number  of  sources,  the  unique  process  that  was 
employed  allowed  the  complicating  situation  of  a  limited  number  of  control  points 
(wells)  to  be  overcome.  The  study  resulted  in  reservoir  models  which  exactly 
honored  the  control  points,  and  which  corroborated  the  field  production  profiles 
of  other  independent  engineering  tests.  The  geostatistical  approach,  however, 


11 


yielded  results  having  much  more  lateral  and  vertical  detail  which  could  be  used  to 
more  effectively  guide  the  placement  of  future  additional  wells  and/or  drilling 
platforms. 

REFERENCES 

Davis,  John  C.  1986.  Statistics  and  data  analysis  in  geology,  2nd  ed.  New  York: 
John  Wiley  &  Sons.  646  p. 

Deutsch,  C.  V.  and  A.  G.  Journel.  1992.  GSLIB-geostatistical  software  library 
and  user's  guide.  New  York:  Oxford  University  Press.  2  diskettes  and  340  p. 

Fielding,  C.  R.  and  R.  C.  Crane.  1987.  An  application  of  statistical  modeling  to 
the  prediction  of  hydrocarbon  recovery  factors  in  fluvial  reservoir  sequences,  in 
Recent  Development  in  Fluvial  Sedimentology  (F.  G.  Ethridge,  R.  M.  Flores, 
and  M.  D.  Harvey,  eds.).  Society  of  Economic  Paleontology  and  Mineralogy, 
Special  Publication  39,  p.  321-327. 

Knutson,  C.  F.  and  C.  R.  Boardman.  1978.  Continuity  and  permeability  develop- 
ment in  the  tight  gas  sands  of  the  eastern  Uinta  Basin,  Utah.  Proceedings  of 
the  Fourth  DOE  Symposium  on  Enhanced  Oil  &  Gas  Recovery  and  Improved 
Drilling  Methods,  v.  2,  p.  E-2/1-24. 

Olea,  Ricardo  A.  1988.  CORRELATOR- An  interactive  computer  system  for 
lithostratigraphic  correlation  of  wireline  logs.  Lawrence,  KS:  Kansas  Geological 
Survey,  Petrophysical  Series  4.  84  p. 

Srivastava,  R.  M.  1992.  Reservoir  characterization  with  probability  field  simula- 
tion. Proceedings  of  the  67th  Annual  Technical  Conference  and  Exhibition  of 
Society  of  Petroleum  Engineers,  SPE  Publication  24753,  p.  927-938. 

BIOGRAPHICAL  SKETCH 

Jeffrey  M.  Yarns  is  a  consulting  mathematical  geologist  and  principal  in  the  firm 
Yarns  &  Associates  in  Denver,  Colorado.  He  holds  a  Ph.D.  in  geology  from  the 
University  of  South  Carolina,  and  specializes  in  applications  of  geostatistics  in 
petroleum  reservoir  characterization,  geoscientific  computer  mapping,  and 
geostatistical  software  development. 

Jeffrey  A.  May  is  an  independent  consulting  geologist  based  in  Littleton, 
Colorado.  He  holds  a  Ph.D.  in  geology  from  Rice  University,  and  specializes  in 
sedimentology,  seismic  interpretation,  and  sequence  stratigraphy. 

Timothy  C.  Coburn  is  senior  statistician  at  the  National  Renewable  Energy 
Laboratory  in  Golden,  Colorado.  He  holds  a  Ph.D.  is  statistics  from  Oklahoma 
State  University,  and  specializes  in  the  analysis  and  modeling  of  energy  and 
environmental  data,  with  particular  emphasis  on  geoscience  investigations. 


12 


An  Overview  of  Stochastic 
Spatial  Simulation 

R.  Mohan  Srivastava1 

Abstract.  —  Though  tools  for  stochastic  spatial  simulation  are 
becoming  increasingly  available  and  accessible,  the  variety  of  dif- 
ferent algorithms  also  forces  users  to  make  conscious  choices  about 
the  methodologies  they  select.  This  overview  paper  presents  a  brief 
summary  of  two  of  the  most  successful  and  popular  approaches,  se- 
quential simulation  and  simulated  annealing,  and  uses  these  as  the 
basis  for  discussing  some  of  the  practical  and  theoretical  issues  sur- 
rounding the  use  of  stochastic  spatial  simulation. 


INTRODUCTION 

In  a  broad  range  of  earth  science  applications,  stochastic  spatial  simulation  is 
rapidly  becoming  a  common  tool  for  engineering  and  decision-making  in  the 
face  of  uncertainty.  In  environmental  applications,  it  is  being  used  as  the  basis 
for  remediation  planning  and  confirmation  sampling,  not  only  at  large  Super- 
fund  sites  but  also  for  small  backyard  contamination  problems.  In  fisheries 
studies,  stochastic  spatial  simulation  is  being  used  by  regulatory  agencies  to 
improve  their  ability  to  manage  fish  stocks  and  to  guarantee  the  stability  of 
the  fisheries  industry.  In  agricultural  studies,  the  same  tools  are  being  used  to 
determine  whether  or  not  policies  designed  to  improve  soil  properties  are  hav- 
ing their  intended  effect.  As  the  interest  in  stochastic  spatial  simulation  grows, 
there  is  an  explosion  in  the  variety  of  simulation  algorithms.  Even  a  compre- 
hensive toolkit  like  the  public  domain  GSLIB  software  (Deutsch  and  Journel, 
1992),  with  its  11  routines  for  simulation,  does  not  cover  the  breadth  of  what 
is  currently  available.  With  simulation  research  thriving  at  several  academic 
institutions,  broad  choice  is  starting  to  become  bewildering  confusion. 

WHAT  DO  WE  WANT  FROM  A  STOCHASTIC  METHOD? 

Stochastic  models  as  art 

Stochastic  models  often  are  not  used  quantitatively  and  serve  only  as  colorful 
wallpaper  for  the  halls  of  research  centers.  Though  such  artwork  is  often  dis- 
missed as  lacking  in  substance,  it  can  play  a  useful  role  in  catalyzing  better 
technical  work.  In  industries  where  there  is  an  endemic  belief  that  all  spatial 
models  should  look  like  contour  maps,  it  is  difficult  to  get  people  to  appreciate 
that  there  is  much  more  complexity  between  the  available  sample  data  than 

1  Geostattstictan,  FSS  International,  Vancouver,  B.C.,  Canada 


13 


traditional  contouring  techniques  can  portray.  Stochastic  models  often  have 
their  greatest  influence  when  they  are  used  to  challenge  a  complacent  belief  in 
the  simplicity  of  spatial  processes.  When  engineers  and  scientists  who  are  fa- 
miliar with  traditional  contouring  techniques  first  see  the  results  of  a  stochastic 
approach,  their  initial  reaction  is  often  one  of  surprise  and  skepticism.  Their 
surprise  is  due  to  the  fact  that  stochastic  models  are  much  more  visually  com- 
plex than  the  models  they  are  familiar  with;  their  skepticism  is  due  to  the  fact 
that  there  are  many  different  renditions,  rather  than  a  unique  "best"  model. 
This  initial  surprise  and  skepticism  often  gives  way  to  curiosity,  however,  as 
they  realize  that  these  alternate  renditions  all  share  a  similar  visual  character 
and  all  somehow  manage  to  honor  the  available  data. 

Once  a  stochastic  model  is  rendered  on  maps  and  cross-sections  and  hung  on 
the  walls  of  a  meeting  room,  it  often  becomes  a  lightning  rod  for  criticism  as 
people  identify  specific  features  that  cannot  possibly  exist.  Their  reasons  vary 
from  geologic  arguments  based  on  depositional  environment  to  geophysical  ar- 
guments based  on  surface  seismic  studies  to  engineering  arguments  based  on 
pressure  tests  to  biologic  arguments  based  simple  principles  of  food  supply. 
In  all  of  these  situations,  the  benefit  of  the  stochastic  model  may  simply  be 
to  focus  the  views  of  a  wide  variety  of  experts  and,  in  so  doing,  to  point  the 
direction  to  an  improved  model.  Whether  this  new  and  clearer  understanding 
is  used  to  improve  the  stochastic  model  or  is  used  to  make  intelligent  adapta- 
tions to  more  traditional  models,  the  initial  stochastic  model  has  still  played 
an  important  role  in  the  evolution  of  an  appropriate  spatial  model. 

Stochastic  models  for  assessing  the  impact  of  uncertainty 

Many  of  the  engineers  and  scientists  who  are  responsible  for  forecasting  the 
future  behavior  of  a  spatial  process,  such  as  fluid  flow  in  a  reservoir  or  contam- 
inant transport  in  an  aquifer,  have  a  solid  understanding  of  the  fact  that  there 
is  always  uncertainty  in  any  spatial  model.  They  know  that  their  forecasts 
are  based  on  a  specific  model  that  is  designated  the  "best"  model,  but  they 
also  want  a  couple  of  other  models,  a  "pessimistic"  one  and  an  "optimistic" 
one,  so  that  they  can  assess  whether  the  decisions  and  policies  that  have  been 
developed  on  the  "best"  model  are  flexible  enough  to  handle  the  uncertainty. 

When  used  for  this  kind  of  study,  a  stochastic  approach  offers  a  variety  of  spa- 
tial models,  each  one  of  which  is  consistent  with  the  available  data.  We  can  sift 
through  these  many  possible  renditions  and  select  one  that  looks  pessimistic 
and  another  that  looks  optimistic. 

Stochastic  models  for  Monte  Carlo  risk  analysis 

Stochastic  modelling  offers  the  hope  of  supporting  a  Monte  Carlo  type  of  risk 
analysis  since  the  various  renditions  it  produces  are  not  only  plausible,  in  the 


14 


sense  that  they  honor  all  the  available  data,  but  they  are  also  intended  to  be 
equally  probable  in  the  sense  that  any  one  of  them  is  as  likely  a  representation 
of  the  true  underlying  spatial  phenomenon  as  any  other  one.  In  such  studies, 
hundreds  of  alternate  models,  if  not  thousands,  are  generated  and  processed  to 
produce  a  distribution  of  possible  values  for  some  critical  engineering  param- 
eter —  breakthrough  time  of  a  contaminant,  for  example,  or  an  economically 
optimal  ore/waste  classification;  these  distributions  are  then  used  to  optimize 
some  decision  by  minimizing  an  objective  function. 

A  critical  aspect  of  this  use  of  stochastic  modelling  is  the  belief  in  a  "space 
of  uncertainty"  and  that  the  stochastic  modelling  technique  can  produce  out- 
comes that  sample  this  space  fairly.  We  know  that  we  cannot  possibly  look  at 
all  possible  outcomes,  but  we  believe  that  we  can  get  a  fair  representation  of 
the  whole  spectrum  of  possibilities.  When  used  for  this  purpose,  we  hope  that 
our  stochastic  modelling  technique  does  not  have  any  systematic  tendency  to 
show  us  optimistic  or  pessimistic  scenarios  since  we  are  going  to  use  all  of  the 
outcomes  as  equally  probable  representations  of  reality. 

While  these  types  of  studies  overlap  somewhat  with  those  described  in  the 
previous  section,  the  important  difference  between  the  two  is  that  Monte  Carlo 
risk  analysis  involves  the  notion  of  a  probability  distribution  while  the  type 
of  study  described  in  the  previous  section  does  not.  The  previous  section 
referred  to  studies  where  models  are  chosen  by  sifting  through  a  large  set 
and  extracting  two  that  seem  to  be  plausible,  but  extreme,  scenarios.  No-one 
particularly  cares  whether  the  "optimistic"  case  is  the  95th  percentile  or  the 
99th  percentile;  it's  just  an  example  of  a  case  that  produces  a  very  optimistic 
result  while  still  honoring  the  same  basic  data.  In  Monte  Carlo  risk  analysis, 
however,  we  depend  on  the  notion  of  a  complete  probability  distribution  of 
possible  outcomes  and  on  the  hope  that  our  stochastic  approach  is  producing 
outcomes  that  fairly  represent  the  entire  space. 

Stochastic  models  for  honoring  spatial  variation 

Though  stochastic  techniques  can  produce  many  outcomes,  stochastic  mod- 
elling studies  often  use  a  single  outcome  as  the  basis  for  forecasting  and  plan- 
ning. When  used  in  this  way,  spatial  simulation  techniques  are  attractive  not 
so  much  for  their  ability  to  generate  many  outcomes  but  rather  for  their  ability 
to  produce  outcomes  that  have  realistic  spatial  variation. 

Over  the  past  decade  it  has  become  increasingly  clear  that  forecasts  of  behavior 
of  spatial  processes  are  more  accurate  when  based  on  models  that  reflect  the 
actual  spatial  heterogeneity  of  the  phenomenon  under  study.  In  the  petroleum 
indsutry,  there  are  countless  examples  of  reservoir  performance  predictions 
whose  failure  is  due  to  the  use  of  overly  simplistic  models;  similar  disasters 
can  be  found  in  environmental  remediation  plans  that  were  based  on  smooth 


15 


contour  maps  that  could  not  possibly  be  used  to  assess  the  effect  of  unexpected 
"hot  spots" .  Most  traditional  methods  for  modelling  spatial  processes  end  up 
with  a  model  that  is  too  smooth  and  continuous,  gently  undulating  from  one 
sample  location  to  the  next  rather  than  showing  the  spatial  variation  that 
is  known  to  exist  between  samples.  Such  smoothness  often  leads  to  biased 
predictions  and  poor  decisions  —  contaminants  escape  from  their  repositories 
much  quicker  than  expected,  for  example,  or  the  actual  volumes  of  remediated 
soil  end  up  being  much  larger  than  originally  forecast. 

Though  using  a  stochastic  method  to  produce  a  single  outcome  is  often  viewed 
with  disdain  by  those  who  prefer  to  pump  hundreds  of  possible  outcomes 
through  Monte  Carlo  risk  analysis,  it  can  be  argued  that  even  a  single  outcome 
from  a  stochastic  approach  is  a  better  basis  for  prediction  and  decision-making 
than  a  single  outcome  from  a  technique  that  does  not  honor  spatial  variation. 

TWO  COMMON  STOCHASTIC  METHODS 
Sequential  simulation 

The  family  of  "sequential"  procedures  all  make  use  of  the  same  basic  algorithm: 

1.  Choose  at  random  a  node  where  no  simulated  value  exists. 

2.  Estimate  the  local  conditional  probability  distribution  (lcpd). 

3.  Draw  at  random  a  single  value  from  the  lcpd. 

4.  Include  the  newly  simulated  value  in  the  set  of  conditioning  data. 

5.  Repeat  steps  1  through  4  until  all  grid  nodes  have  a  simulated  value. 


(a)  CONDITIONING  DATA        (b)  RANDOM  LOCATION 


(c)  LCPD  AT  ® 


(d)  SIMULATED  VALUE 


17.1 

26.2  A 

• 

•  f 

11.2 

21.9 

• 

• 

17.1 

• 

26.2  ^ 

® 

11.2 

21.9 

• 

• 

10  20 
POROSITY 


Figure  1:  Sequential  simulation  of  porosities  in  an  aquifer. 


The  only  significant  difference  between  the  various  sequential  procedures  is 
estimation  of  the  lcpd;  any  technique  that  estimates  the  lcpd  can  be  used  to 
drive  sequential  simulation.  Multigaussian  kriging,  for  example,  produces  an 
estimate  of  the  lcpd  by  assuming  that  it  follows  the  normal  distribution  and 
estimating  its  mean  and  standard  deviation;  this  was  the  approach  used  in 
the  diagram  in  Figure  1.  Indicator  kriging  is  an  example  of  another  technique 
that  could  be  used  to  estimate  the  lcpd.  With  this  procedure,  no  assumption 
is  made  about  the  shape  of  the  distribution  and  the  lcps  is  contructed  by 
directly  estimating  the  probability  of  being  below  a  series  of  thresholds  or  the 
probability  of  being  within  a  set  of  discrete  classes. 


16 


(a)  INITIAL  GRID 

Note:  Each  pixel  is  20m  Net-to-gross: 
horizontally  by  3m  vertically     Ave.  length: 
Ave.  thickness: 


Actual 

69.7% 
29.78m 
4.53m 


Target 

70% 
60m 
10m 


(b)  ITERATION  NO.  1 

(2.17)  «->  (25.25) 
Statistics  improve, 
keep  this  grid 


Actual  Target 

Net-to-gross:  69.7%  70% 

Ave.  length:  30.00m  60m 

Ave.  thickness:   4.57m  10m 


(c)  ITERATION  NO.  2 

(2.13)  «-» (23.15) 
Statistics  deteriorate, 
go  back  to  last  grid 


Actual  Target 

Net-to-gross:  69.7%  70% 

Ave.  length:  29.34m  60m 

Ave.  thickness:  4.45m  10m 


(d)  ITERATION  NO.  463 

(24,5)  <->  (7,12) 
Statistics  improve, 
keep  this  grid 


Net-to-gross:  69.7% 
Ave.  length:  45.68m 
Ave.  thickness:  6.55m 


Actual  Target 

70% 
60m 
10m 


(e)  ITERATION  NO.  464 

(2,5)  <->  (8,22) 
Statistics  deteriorate, 
go  back  to  last  grid 


Actual  Target 


Net-to-gross: 
Ave.  length: 
Ave.  thickness: 


69.7% 
45.46m 
6.51m 


70% 
60m 
10m 


(f)  ITERATION  NO.  1987 

(25,14) «-» (14,12) 
Statistics  improve, 
keep  this  grid 


Net-to-gross 
Ave.  length 
Ave.  thickness 


Actual 

69.7% 
60.00m 
9.88m 


Target 

70% 
60m 
10m 


Figure  2:  Simulated  annealing  for  a  simple  sand/shale  simulation. 


Simulated  annealing 

In  simulated  annealing  the  spatial  model  is  constructed  by  iterative  trial-and- 
error.  Figure  2  shows  a  simplified  annealing  procedure  for  the  problem  of 
producing  a  sand/shale  model  that  has  a  70%  sand,  an  average  shale  length 
of  60  meters  and  an  average  thickness  of  10  meters.  In  Figure  2a  the  pixels 
in  the  model  are  randomly  initialized  with  sands  and  shales  in  the  correct 
global  proportion.  The  random  assignment  of  the  sands  and  shales  causes  the 
average  shale  length  and  thickness  to  be  too  short.  Figure  2b  shows  what 
happens  after  we  perturb  the  initial  model  by  swapping  a  sand  pixel  with  a 
shale  pixel.  Such  a  swap  will  not  affect  the  global  proportions  of  sand  and 
shale:  it  does,  however,  make  a  tiny  improvement  in  both  the  average  shale 
length  and  thickness.  Figure  2c  shows  the  grid  after  a  second  swap.  We're  not 


17 


quite  as  lucky  with  this  second  swap  since  it  makes  the  average  shale  length 
and  thickness  a  little  worse,  so  we  reject  this  second  swap,  go  back  to  the 
previous  grid  shown  in  Figure  2b  and  try  again. 

This  swapping  is  repeated  many  more  times  and,  after  each  swap,  the  average 
shale  length  and  thickness  are  checked  to  see  if  we  are  any  closer  to  the  target 
values.  If  a  particular  swap  does  happen  to  get  us  closer  to  the  length  and 
thickness  statistics  we  want  then  we  keep  the  swap;  on  the  other  hand,  if 
the  swap  causes  the  image  to  deteriorate  (in  the  sense  that  the  length  and 
thickness  statistics  are  further  from  our  target)  then  we  undo  the  swap,  go 
back  to  the  previous  grid  and  try  again. 

Figure  2d  shows  the  grid  after  the  463rd  swap,  which  happens  to  be  a  good 
one.  At  this  point,  66  of  the  attempted  463  swaps  were  good  and  were  kept, 
while  the  remaining  397  were  bad  and  were  not  kept.  Figure  2e  shows  the 
effect  of  the  464th  swap,  a  bad  one  that  is  not  kept  because  it  causes  the 
statistics  to  deteriorate.  Figure  2f  shows  what  the  grid  looks  like  after  the 
1987th  swap.  At  this  point  there  were  108  good  swaps  and  1879  bad  ones,  and 
the  shale  length  and  thickness  statistics  are  very  close  to  our  target  values. 

The  key  to  simulated  annealing  is  minimizing  the  deviation  between  the  grid 
statistics  and  target  values;  the  function  that  describes  this  deviation  is  usu- 
ally called  the  "energy"  or  "objective"  function.  In  the  example  in  Figure  2, 
where  we  were  trying  to  match  the  average  length  and  thickness  of  the  shales, 
the  energy  function  was  simply  the  sum  of  the  absolute  differences  between 
the  actual  average  dimensions  and  our  target  values:  E  =  | Actual  average 
length— 60 j  +  |  Actual  average  thickness— 10 1.  The  annealing  procedure  tries  to 
drive  the  value  of  E  to  zero;  decreases  in  energy  are  good,  increases  are  bad. 

Though  Figure  2  displays  some  of  the  key  concepts  in  simulated  annealing,  it 
is  a  very  simplified  version  and  does  not  show  some  of  the  important  aspects  of 
practical  implementations  of  simulated  annealing.  Most  simulated  annealing 
procedures  do  not  recalculate  the  energy  after  each  swap.  What  they  do 
instead  is  perform  several  swaps  before  recalculating  the  energy.  Initially, 
when  the  energy  is  high  (the  statistics  of  the  grid  are  far  away  from  the  target 
values)  many  swaps  are  performed  before  the  energy  is  recalculated;  as  the 
energy  decreases  and  the  grid  statistics  converge  to  their  desired  target  values, 
the  number  of  swaps  between  recalculations  is  decreased. 

The  second  important  difference  between  most  practical  implementations  of 
annealing  and  the  simplified  example  in  Figure  2  is  tolerance  of  lack  of  progress. 
In  the  example  shown  in  Figure  2,  any  swap  that  increased  the  energy  (moved 
us  further  away  from  our  target  statistics)  was  immediately  undone.  In  the 
common  practical  implementations  of  simulated  annealing,  some  increases  in 
energy  are  tolerated.  The  chance  of  allowing  the  energy  to  increase  is  related 
to  two  factors:  the  magnitude  of  the  increase  and  the  number  of  iterations.  In- 


18 


creases  in  energy  are  more  tolerated  early  on  in  the  procedure;  small  increases 
are  always  more  tolerated  than  large  ones. 

A  third  difference  between  the  example  given  here  and  most  practical  imple- 
mentations is  that  the  example  in  Figure  2  does  not  have  any  conditioning 
information  to  honor.  Conditioning  information  is  quite  easily  honored  in 
practice  by  setting  the  appropriate  pixels  to  the  values  observed  and  then  en- 
suring that  these  conditioning  pixels  are  never  swapped  later  in  the  procedure. 

The  other  important  difference  between  most  practical  implementations  of 
simulated  annealing  and  the  simplified  example  in  Figure  2  is  that  the  energy 
function  is  usually  much  more  complex  than  the  simple  one  used  here.  Simu- 
lated annealing  is  a  very  flexible  approach  that  can  produce  stochastic  models 
that  honor  many  different  kinds  of  information;  the  paper  by  Kelkar  and  Shibli 
in  this  volume  provides  an  example  of  the  use  of  annealing  for  honoring  the 
fractal  dimension.  For  any  desired  features  or  properties  that  we  can  sum- 
marize numerically,  we  can  use  simulated  annealing  to  honor  these  features 
by  including  appropriate  terms  in  the  energy  function.  In  Figure  2,  for  ex- 
ample, we  wanted  to  honor  both  the  average  shale  thickness  and  the  average 
shale  length  so  our  energy  function  had  two  components.  In  more  interesting 
practical  examples,  the  energy  function  might  have  several  more  components. 
For  example,  we  could  have  one  component  that  describes  how  close  the  grid's 
variogram  comes  to  some  target  variogram,  a  second  component  that  describes 
how  similar  the  synthetic  seismic  response  of  the  grid  comes  to  some  actual 
seismic  data,  and  a  third  component  that  describes  how  close  a  simulated 
tracer  test  on  the  grid  comes  to  observed  tracer  test  results. 

ADVANTAGES  AND  DISADVANTAGES 
Stochastic  models  as  art 

If  a  stochastic  model  is  being  used  to  visually  catalyze  more  critical  thinking, 
then  one  of  the  primary  criteria  in  choosing  a  method  has  to  be  the  visual  ap- 
peal of  the  final  result.  For  stochastic  modelling  of  geological  objects,  like  sand 
channels,  pixel-based  techniques,  like  the  sequential  procedures  discussed  ear- 
lier, often  tend  to  produce  geological  objects  that  are  too  broken  up  around  the 
edges.  Simulated  annealing  is  able  to  better  preserve  the  sharp  edged  features 
that  many  geologists  expect  to  see.  There  are  also  object-based  techniques, 
also  often  referred  to  as  "boolean"  methods,  whose  visual  appeal  is  hard  to 
beat  because  they  deal  with  entire  objects  rather  than  elementary  pixels. 

For  stochastic  modelling  of  continuous  variables,  such  as  porosities  or  con- 
taminant concentrations,  a  boolean  approach  is  not  going  to  completely  do 
the  trick.  Such  an  approach  is  usually  useful  as  a  first  step,  but  the  art- 
work is  unconvincing  if  continuous  properties  within  each  population  are  as- 
signed randomly.  The  images  look  more  plausible  when  the  properties  within 


19 


a  given  population  have  a  prescribed  pattern  of  spatial  continuity.  With  most 
of  the  techniques  for  modelling  continuous  variables,  the  spatial  continuity  is 
described  through  a  variogram  model.  Though  experimental  variograms  typ- 
ically show  a  nugget  effect,  experience  has  shown  that  stochastic  models  of 
continuous  properties  have  more  visual  appeal  if  the  variogram  model  used  in 
the  stochastic  simulation  procedure  has  no  nugget  effect. 

For  stochastic  modelling  of  a  very  continuous  property,  such  as  the  thick- 
ness of  an  aquifer,  sequential  techniques  tend  not  to  work  well.  The  same 
variogram  models  that  describe  very  continuous  surfaces  —  the  gaussian  var- 
iogram model,  for  example,  or  a  power  model  with  an  exponent  close  to  2 
—  also  produce  chaotic  weighting  schemes  when  some  of  the  nearby  samples 
fall  directly  behind  closer  samples.  Since  sequential  procedures  include  each 
simulated  value  as  hard  data  that  must  be  taken  into  account  when  future 
nodes  are  simulated,  and  since  they  are  typically  performed  on  a  regular  grid, 
chaotic  weighting  schemes  are  virtually  guaranteed  whenever  we  are  trying  to 
simulate  a  very  continuous  surface. 

Most  of  the  stochastic  modelling  techniques  for  continuous  variables  give  us 
the  ability  to  honor  a  variogram,  or  a  set  of  indicator  variograms.  If  we  are 
trying  to  produce  stochastic  artwork  that  shows  specific  sedimentary  features, 
such  as  a  succession  of  fining-upwards  sequences,  or  a  fabric  of  cross-cutting 
features,  neither  of  which  is  well-described  by  a  variogram,  then  annealing  is 
often  the  most  useful  approach.  As  long  as  we  can  find  a  way  of  numerically 
summarizing  the  sedimentary  features  that  we  would  like  to  see,  then  we  can 
include  appropriate  terms  in  our  energy  function  for  annealing. 

Stochastic  models  for  assessing  the  impact  of  uncertainty 

When  the  goal  of  a  study  is  to  produce  a  handful  of  models  that  capture 
a  couple  of  extreme  scenarios  as  well  as  some  intermediate  ones,  one  of  the 
practical  issues  that  comes  to  the  forefront  is  computational  speed.  This  is 
especially  true  because  it  is  usually  not  possible  to  steer  a  stochastic  model  to 
an  optimistic  realization  or  to  a  pessimistic  one;  the  only  way  of  finding  such 
extreme  scenarios  is  to  produce  many  realizations  and  sift  through  them  all. 

It  is  difficult  to  get  an  objective  ranking  of  the  speed  of  stochastic  modelling 
algorithms  since  their  authors  are  tempted  to  brag  that  their  own  method 
is  lightning  fast.  Though  all  methods  are  workable  in  practice,  some  require 
several  days  of  run-time  on  fast  computers  to  produce  a  single  realization 
despite  their  author's  enthusiastic  claims  of  speed.  Such  procedures  cannot  be 
a  practical  basis  for  producing  many  realizations. 

As  a  general  rule,  sequential  procedures  are  slow  because  they  have  to  search 
through  an  ever-increasing  set  of  hard  data  to  find  the  nearby  samples  that 
are  relevant  for  the  estimation  of  the  local  conditional  probability  distribution. 


20 


The  notable  exceptions  to  this  are  the  sequential  simulation  procedures  based 
on  Markov  random  fields.  These  sequential  procedures  considerably  accelerate 
the  search  for  previously  simulated  data  by  using  only  the  closest  ring  of  data; 
some  of  them  accomplish  even  greater  speed  by  precalculating  the  kriging 
weights  that  will  eventually  be  needed  during  the  sequential  simulation. 

Annealing  is  a  highly  variable  procedure  in  terms  of  its  run-time.  For  certain 
problems,  annealing  is  very  rapid;  for  others  it  is  abysmally  slow.  The  key 
to  rapid  annealing  is  the  ability  to  update  rather  than  recalculate  the  energy 
function.  If  a  global  statistic  has  to  be  recalculated  from  scratch  every  time  two 
pixels  are  swapped,  then  annealing  is  not  likely  to  be  speedy;  if  there  is  some 
convenient  trick  to  updating  the  previous  calculation  rather  than  recalculating 
from  scratch,  then  annealing  can  be  very  efficient. 

Stochastic  models  for  Monte  Carlo  risk  analysis 

For  studies  that  involve  Monte  Carlo  risk  analysis,  there  are  two  criteria  that 
need  to  be  considered  when  selecting  a  stochastic  modelling  technique.  The 
first  is  computational  speed;  the  second  is  the  issue  of  whether  the  stochastic 
method  is  fairly  sampling  the  space  of  uncertainty.  The  issue  of  the  space  of 
uncertainty  is  one  of  the  thorniest  theoretical  problems  for  stochastic  mod- 
elling. There  are  some  who  take  the  view  that  the  space  of  uncertainty  must 
be  theoretically  defined  outside  the  context  of  the  algorithm  itself  so  that  the 
user  does  have  some  hope  of  checking  that  a  family  of  realizations  represents  a 
fair  sampling  of  the  space  of  uncertainty.  Others,  however,  take  the  view  that 
the  space  of  uncertainty  can  only  be  defined  through  the  algorithm  and  that  it 
is,  by  definition,  the  family  of  realizations  that  is  produced  when  all  possible 
random  number  seeds  are  fed  to  the  algorithm. 

This  issue  of  the  equiprobability  of  the  realizations  and  the  fair  sampling  of 
the  space  of  uncertainty  comes  to  the  forefront  when  annealing  is  used  as 
a  post-processor  of  the  results  from  some  other  technique.  Though  there  is 
some  theory  that  backs  up  the  claim  that  annealing  fairly  samples  the  space  of 
uncertainty,  this  theory  is  entirely  based  on  the  assumption  that  the  starting 
grid  is  random  (or  that  the  initial  energy  is  very  high).  Though  annealing  offers 
tremendous  flexibility  in  incorporating  and  honoring  widely  different  types  of 
information,  there  remains  an  awkward  ambiguity  about  its  implicit  space  of 
uncertainty  and  whether  this  space  is  being  fairly  sampled. 

Stochastic  models  for  honoring  spatial  variation 

If  the  reason  for  producing  a  stochastic  model  is  to  respect  critical  hetero- 
geneities, then  the  choice  of  a  stochastic  modelling  method  should  take  into 
account  the  type  of  data  that  exists  on  heterogeneities.  Annealing  is  a  pow- 
erful technique  for  accommodating  information  on  heterogeneities  that  comes 


21 


through  complex,  indirect  or  time-varying  measurements,  such  as  pressure 
transient  tests  in  aquifers.  Annealing  has  been  used,  for  example,  to  directly 
honor  tracer  test  results;  though  the  procedure  is  very  slow  (because  the  tracer 
test  response  cannot  be  readily  updated  when  pixels  are  swapped),  the  result- 
ing scenarios  are  tremendously  realistic  since  they  capture  the  connectivity  of 
the  extreme  values  that  is  implicit  in  the  actual  tracer  test  observations. 

CONCLUSIONS 

Though  this  paper  has  tended  to  discuss  stochastic  simulation  methods  in 
isolation,  many  practical  problems  call  for  a  combination  of  appropriate  simu- 
lation algorithms.  It  is  unfortunate  that  much  of  the  technical  literature  and 
many  of  the  commercial  software  packages  for  stochastic  spatial  modelling 
promote  a  single  method  as  better  than  all  others.  There  is  no  stochastic 
modelling  method  that  is  universally  best  for  all  possible  spatial  phenomena. 
As  stochastic  modelling  becomes  more  accepted  in  the  earth  sciences  and  as 
new  stochastic  modelling  techniques  are  developed,  it  is  certain  that  the  most 
successful  applications  will  be  those  that  view  the  growing  assortment  of  meth- 
ods as  a  complementary  toolkit  rather  than  as  a  set  of  competing  methods. 

REFERENCES 

Deutsch,  C.V.,  and  Journel,  A.G.,  1992,  GSLIB:  Geostatistical  Software  Li- 
brary and  User's  Guide,  Oxford  University  Press,  340  p. 

Yarus,  J.,  and  Chambers,  R.,  (eds.),  1994,  Stochastic  Modeling  and  Geostatis- 
tics,  Practical  Applications  and  Case  Histories,  American  Association  of 
Petroleum  Geologists. 

BIOGRAPHICAL  SKETCH 

R.  Mohan  Srivastava,  is  a  geostatistical  consultant  with  fifteen  years  of  ex- 
perience in  the  practice  of  geostatistics.  He  is  an  author  of  An  Introduction 
to  Applied  Geostatistics,  the  major  introductory  textbook  on  the  practice  of 
geostatistics,  and  of  more  than  thirty  technical  articles  and  reports  on  the 
theory  and  practice  of  geostatistics.  In  addition  to  using  geostatistics  for  both 
exploration  and  development  projects  with  the  oil  and  mining  industries,  he 
has  also  applied  geostatistics  to  a  wide  variety  of  environmental  and  agricul- 
tural problems.  He  is  one  of  the  founding  partners  of  FSS  International,  a 
consulting  group  that  specializes  in  geostatistics,  and  currently  manages  their 
Canadian  operations  from  his  office  in  Vancouver. 


22 


Choosing  and  Using  Simulation 

Algorithms 

Donald  E.  Myers1 


Abstract.— In  geostatistics  simulation  means  generating  another  "reality",  i.e, 
another  realization  of  a  random  function.  Usually  the  random  function  is  only 
characterized  in  terms  of  the  first  two  moments,  the  mean  and  a  spatial 
correlation  function  such  as  the  covariance  or  variogram.  Interpolation 
methods  allow  predicting  or  estimating  values  at  non  data  points  of  the  (same?) 
realization.  Simulation  algorithms  generate  additional  realizations,  often 
conditioned  to  the  original  data. 

Originally  simulation  was  primarily  of  interest  as  a  planning  tool  but  now 
some  are  using  simulation  in  lieu  of  interpolation.  Interpolation  methods 
smooth  the  data,  but  simulation  algorithms  enhance  the  variation.  In  some 
applications  the  smoothing  produces  results  that  are  thought  to  be  less  realistic 
than  simulation. 

There  are  multiple  simulation  algorithms  available,  the  user  must  choose 
between  them.  Simulated  results  are  often  used  as  an  input  for  other  modeling 
steps.  There  are  both  practical  and  theoretical  aspects  to  the  problem  of 
choosing  and  using  a  simulation  algorithm.  Multiple  simulation  algorithms 
will  be  briefly  reviewed  and  properties  discussed  along  with  typical 
applications.  Practical  and  theoretical  aspects  of  equivalence  will  be  discussed 
including  the  sense  in  which  properties  of  the  random  function  are  reproduced 


INTRODUCTION 

As  the  term  is  commonly  used  in  statistics  and  probability,  simulation  means 
generating  a  set  of  numbers  using  a  given  probability  distribution,  standard  limit 
theorems  ensure  that  the  empirical  distribution  (and  its  characteristics)  will 
approximate  those  of  the  original  distribution.  This  easily  extends  to  vector  valued 
random  variables  and  a  joint  distribution.  In  either  case  the  distribution  function  is 
an  essential  tool.  The  other  essential  tool  is  a  random  number  generator  for  the 
Uniform  Distribution  on  an  interval. 

In  geostatistics  the  underlying  model  is  a  random  function  defined  in  1,2  or  3 
dimensional  space,i.e.,  a  random  variable  whose  values  are  functions  rather  than 


Professor  of  Mathematics,  University  of  Arizona,  Tucson,  AZ  85721 


23 


numbers.  Simulation  means  generating  an  alternative  "reality",  i.e.,  another  one  of 
the  functions  hence  preserving  certain  characteristics  of  the  random  function.  Since 
simulation  is  a  numerical  technique  the  realization  is  only  known  at  a  finite  number 
of  points.  In  practice  it  is  more  like  the  problem  of  simulating  a  vector  valued 
random  variable  except  that  the  joint  distribution  function  is  seldom  known.  This 
has  led  to  the  development  of  a  number  of  algorithms  with  differing  theoretical  and 
practical  characteristics. A  user  must  choose  between  algorithms  and  the  choice  of 
an  algorithm  is  driven  in  part  by  the  ultimate  objective,  i.e.,  the  intended  use  of  the 
simulated  values  and  over  time  these  have  changed. 

OBJECTIVES  AND  MODELS 

Early  applications  of  geostatistics  were  principally  in  mining  and  hydrology, 
objectives  for  simulation  were  mainly  related  to  these.  In  mining  simulation  was 
proposed  as  a  planning  tool,  Journel  (1974).  In  hydrology  simulated  values  of 
hydrological  parameters  such  as  hydraulic  conductivity  are  used  as  an  input  for  flow 
models.  Simulation  can  be  used  to  design  sampling  plans  by  generating  multiple 
realizations,  optimizing  the  plan  for  each  and  then  optimizing  between  plans. 
Because  interpolation  smoothes  the  data,  interpolated  surfaces  are  sometimes 
thought  to  be  unrealistic  and  simulation  is  used  instead,  Dalle-Gupta  et  al  (1995). 
Simulated  data  has  also  been  used  to  study  the  behavior  of  various  variogram 
estimators.  Consider  a  random  function  model 

Z(x)  =  m(x)  +  Y(x)  [1] 

where  Y(x)  is  a  zero  mean  random  function  with  some  form  of  stationarity  assumed, 
i.e.,  second  order  or  intrinsic  of  order  k,  and  m(x)  is  the  expected  value  of  Z(x), 
usually  assumed  to  be  a  linear  combination  of  known  basis  functions  such  as 
monomials  in  the  position  coordinates.  Y(x)  is  characterized  by  a  spatial  structure 
function  such  as  a  covariance,  variogram  or  generalized  covariance.  In  addition  Y(x) 
is  usually  assumed  to  have  a  known  univariate  stationary  distribution. 

A  QUICK  OVERVIEW  OF  COMMON  ALGORITHMS 

Turning  Bands 

Let  X(t)  be  a  random  function  defined  in  1  space,  s  a  random  unit  vector  in  n- 
space  and  x  a  point  in  n-space  then  the  scalar  product  <x,s>  is  the  projection  of  the 
point  x  onto  the  direction  s.  By  integrating  with  respect  to  the  random  direction  s, 
one  obtains  a  random  function  Z(x)  defined  in  n-space.  This  provides  a  relationship 
between  the  covariance  function  Ci(r)  for  X(t)  and  the  covariance  Cn(r)  of  Z(x). 
Given  a  covariance  for  n=3  one  can  easily  solve  for  the  corresponding  covariance 
in  1 -space.  Then  by  writing  the  1 -dimensional  covariance  as  a  convolution  one 


24 


obtains  a  function  that  can  be  used  in  a  Box-Jenkins,  i.e.,  moving  average,  simulation 
algorithm  (in  1 -space).  To  discretize  the  algorithm  the  user  must  make  choices, 
first,  one  can  not  use  an  infinite  number  of  directions.  They  should  be  equally  spaced 
and  in  3-dimensions  geometry  intervenes  to  show  that  15  is  maximum  number. 
Theoretically  the  Moving  Average  simulation  is  an  integral  and  must  be 
approximated  by  an  infinite  series  but  only  a  finite  number  of  terms  are  used. 
Instead  of  simulating  a  value  at  each  point  on  a  line,  one  simulates  a  value  for  each 
of  a  set  of  disjoint  intervals.  When  the  lines  are  rotated  in  different  directions  we 
obtain  the  "turning  bands". .  Given  a  point  x  then  one  determines  the  interval  on  each 
direction  that  the  point  projects  into  and  looks  up  the  simulated  value  for  that 
interval.  The  simulated  value  for  the  point  x  is  then  simply  a  sum  of  those  values. 

While  the  integral  equation  linking  the  1 -dimensional  covariance  and  the  n- 
dimensional  covariance  is  easily  solved  for  n=3,  it  is  not  so  simple  for  n=2.  This 
led  to  the  use  of  spectral  methods,  see  Mantoglou  and  Wilson  ( 1 982)  for  applications 
in  hydrology  and  Booker  (19  85)  for  solutions  of  the  integral  equation  when  n=2. 
There  is  another  important  difference  between  the  cases  for  n=2  and  n=3.  For  any 
even  integer  it  is  easy  to  choose  equally  spaced  directions  in  2-space  (equally  spaced 
points  on  the  unit  circle).  There  is  no  upper  bound  on  the  number  of  equally  spaced 
directions. 

The  Turning  Bands  algorithm  will  reproduce  the  covariance  function  in  an 
"average"  sense,  i.e.,  if  one  generates  multiple  realizations  then  averages  the  sample 
covariance  functions  (or  sample  variograms)  the  average  will  approximate  the 
original  covariance  (variogram). 

L-U  Decomposition 

Given  locations  X!,....,xm  the  covariance  matrix  of  the  random  variables 
Z(xi),....,Z(Xn,)  is  positive  definite  and  may  be  "factored"  in  at  least  two  ways. The 
Cholesky  decomposition  produces  two  matrices  L,  U  (which  are  respectively  lower 
triangular  and  upper  triangular)  such  that  LT  =  U.  Now  suppose  that  Yu....,Yn  are 
uncorrected,  mean  zero,  variance  one  random  variables.  The  product  UY  is  a 
collection  of  random  variables  whose  covariance  matrix  is  the  original.  Although 
the  technique  was  well-known  in  statistics,  the  application  and  extension  to  spatial 
data  is  found  in  Davis  (1987)  where  it  is  also  shown  that  the  conditioning  is  easily 
incorporated.  There  are  at  least  two  difficulties;  most  algorithms  for  finding  the  L- 
U  decomposition  require  that  the  entire  matrix  be  in  memory  at  one  time  and  this 
limits  the  size  of  the  grid  to  about  1000  points.  Alabert  (1987)  has  given  a  modified 
form  of  the  algorithm  allowing  the  use  of  a  moving  neighborhood  for  the 
conditioning  and  hence  the  use  of  multiple  covariance  matrices  instead  of  one  large 
matrix.  Since  the  matrix  "slides"  over  the  grid  one  must  still  consider  the  possible 
discrepancies  that  might  be  generated  by  having  a  grid  location  occur  in  multiple 


25 


matrices. 

Although  it  does  not  completely  solve  the  problem,  Davis  (1987)  also  proposed 
the  use  of  the  square  root  of  the  covariance  matrix  which  is  then  approximated  by 
a  polynomial  expansion  which  only  requires  computing  powers  of  the  matrix  also 
see  Dietrich  and  Newsam  (1995).  Myers  (1989)  extended  both  of  these  algorithms 
to  the  vector  case  including  the  undersampled  case  for  conditioning. 

One  advantage  of  the  L-U  or  square  root  algorithms  is  that  they  are  dimension 
independent  but  both  suffer  the  same  difficulty  with  respect  to  the  univariate 
distribution  as  does  the  Turning  Bands  method,  namely  that  the  simulated  values 
are  obtained  as  linear  combinations  of  uncorrected  random  variables.  Hence  a 
transformation  to  normality  is  needed  along  with  a  reverse  transformation.  There 
is  a  parallel  processing  version  of  the  Cholesky  algorithm  which  would  allow  the 
use  of  larger  covariance  matrices,  i.e.,  larger  grids,  but  it  seems  not  to  have  been 
used  in  geostatistical  applications  as  yet. 

Sequential  Gaussian 

It  is  well-known  that  in  the  case  of  multivariate  normality  the  Simple  Kriging 
estimator  is  the  conditional  mean,  i.e.,  the  mean  of  the  conditional  distribution,  and 
the  Kriging  variance  is  the  conditional  variance.  The  Sequential  Gaussian  simulation 
algorithm  is  based  on  this  property.  Given  a  set  of  data  locations  and  a  grid  point  to 
be  simulated,  one  computes  the  Simple  Kriging  estimated  value  for  the  grid  point 
and  the  kriging  variance.  These  are  used  in  the  conditional  normal  distribution  which 
is  then  "sampled".  Having  generated  a  value  at  the  grid  point,  this  location  and  value 
are  considered  to  be  a  part  of  the  data  set  and  process  is  repeated  at  a  new  grid  point. 
Since  it  is  straightforward  to  use  a  moving  search  neighborhood  for  the  simple 
kriging  it  is  not  necessary  to  use  the  entire  "data"  set  for  each  grid  point  but  rather 
only  those  that  are  close.  The  user  must  then  select  the  parameters  for  the  search 
neighborhood  and  also  select  the  order  in  which  grid  points  are  simulated  since 
points  selected  earlier  may  be  used  in  the  subsequent  simulations. 

Since  normality  is  explicitly  used  in  the  algorithm  it  is  then  necessary  to  transform 
the  data  and  to  re-transform  the  simulated  values.  As  with  the  Turning  Bands  and  L- 
U  decomposition  methods,  the  covariance  used  pertains  to  and  is  modeled  from  the 
transformed  data.  In  an  average  sense  it  can  be  shown  that  any  randomly  selected 
sequence  of  grid  points  produces  the  same  set  of  realizations. 

Simulated  Annealing 

The  title  is  somewhat  mis-leading  since  Simulated  Annealing  is  really  an 
optimization  method.  A  realization  is  generated  by  minimizing  the  "difference" 


26 


between  the  sample  variogram  (or  sample  covariance)  and  the  theoretical  model. 
First  the  data  is  used  to  model  a  probability  distribution,  then  using  this  distribution 
and  a  random  number  generator,  a  value  is  simulated  at  each  grid  location 
(independently).  Although  this  will  ensure  that  the  univariate  distribution  is 
preserved  it  would  in  general  result  in  a  pure  nugget  effect  variogram.  Next  select 
at  random  two  locations,  interchange  the  values  at  these  two  points  and  compute  the 
sample  variogram.  If  the  new  sample  variogram  is  closer  the  interchange  is  retained, 
otherwise  not.  New  locations  are  selected  and  the  process  repeated.  A  "temperature" 
function  is  used  to  ensure  that  the  iteration  stops.  For  a  brief  introduction  see 
Deutsch(1995),  Deutsch  and  Cockerham  (1994). 

VARIATIONS  AND  EXTENSIONS 

The  four  methods  described  above  essentially  assume  that  the  random  function  is 
continuous  valued.  At  least  two  variations  occur  on  this  theme,  coded  data  obtained 
by  for  example  an  indicator  transform  applied  to  continuous  data,  and  categorical 
data.  One  might  simulate  indicator  transform  data  for  reasons  that  are  essentially  the 
same  as  those  related  to  the  use  of  indicator  kriging  as  opposed  to  kriging  of  the 
original  variable.  Journel  and  Isaaks  (1984)  use  an  indicator  transform  to  separate  the 
data  corresponding  to  a  high  grade  mineralization  from  that  corresponding  to  a  low 
grade  mineralization.  Bierkens  and  Weerts  (1994)  apply  indicator  coding  to 
categorical  data,  e.g.,  soil  texture  classes.  In  this  case  it  is  necessary  to  assign  an 
(arbitrary)  ordering  to  the  texture  classes.  The  use  of  indicators  essentially 
corresponds  to  assigning  a  probability  distribution  to  each  point,  i.e.,  one  is 
simulating  a  probability  distribution  at  each  point.  This  distribution  is  conditioned 
on  the  "data".  As  in  indicator  or  probability  kriging,  if  the  values  are  kriged  or 
simulated  for  each  cut-off  separately  then  the  order  relations  may  not  be  satisfied  and 
adjustments  are  needed. 

SOME  QUESTIONS 

Usually  one  desires  that  the  univariate  marginal  distribution  be  preserved  and  this 
also  creates  a  problem.  First  of  all  one  must  distinguish  between  a  theoretical 
distribution  for  the  random  function  model  and  the  spatial  frequency  distribution 
of  a  particular  realization  although  the  latter  is  what  is  usually  computed.  Since  the 
simulated  value  at  any  point  on  the  grid  is  obtained  as  a  sum  of  independent^. e., 
uncorrected,  random  numbers  one  must  expect  a  form  of  the  Central  Limit  Theorem 
effect.  Actually  one  need  not  resort  to  the  Central  Limit  Theorem,  there  are  only  a 
few  distribution  types  for  which  the  sum  of  independent  random  variables  has  the 
same  distribution  type  as  the  summands.  These  include  the  Normal  and  the  Gamma, 
hence  it  is  common  to  utilize  a  transformation  to  transform  the  original  data 
frequency  distribution  to  Normal  and  then  to  re-transform  the  simulated  values  back. 
Since  the  original  data  set  is  finite,  the  empirical  distribution  will  be  discrete  and 


27 


hence  some  smoothing  is  needed  to  transform  to  Normal,  the  transformation  is  not 
uniquely  determined  but  requires  some  user  choices.  In  addition  the  covariance  used 
in  the  simulation  step  reflects  the  spatial  structure  of  the  transformed  data  and  in 
general  there  is  no  assurance  that  the  spatial  structure  of  the  original  data  is 
preserved  in  the  process.  A  desire  to  simplify  the  algorithms  gave  rise  to  an  emphasis 
on  the  LU  decomposition  and  sequential  gaussian  methods  both  of  which  use 
normality  more  explicitly.  As  the  need  for  application  to  non-continuous  variables 
and  also  for  simulation  on  much  larger  grids  increased,  other  algorithms  were 
introduced.  The  choice  between  the  algorithms  was  and  is  largely  driven  by  personal 
preference  or  practical  aspects  of  the  application.  Underlying  this  diversity  of 
algorithms  was  an  implicit  but  never  stated  assumption  that  there  was  some  form  of 
equivalence  and  hence  the  difference  was  only  computational.  This  might  be 
interpreted  as  meaning  that  if  one  considers  the  set  of  all  possible  realizations  that 
could  be  generated  (for  a  given  set  of  parameter  assumptions)  was  the  same  for  all 
methods  or  at  least  approximately  so.  Since  in  practice  one  only  generates  a  small 
number  of  realizations,  i.e.,  the  values  at  a  finite  number  of  points  for  each  of  a  small 
number  of  realizations  there  is  another  implicit  assumption,  that  for  the  particular 
purpose  or  objective  the  conclusions  were  essentially  the  same  irrespective  of  which 
small  set  of  realizations  is  used.  Neither  of  these  implicit  assumptions  has  really  been 
tested  or  even  considered,  most  users  do  not  use  multiple  algorithms  and  make 
comparisons  nor  do  they  generate  multiple  finite  sets  of  realizations  to  compare 
between  the  sets. 

Characterizing  or  defining  "equivalence"  is  non-trivial.  One  very  strong  perception 
would  be  that  if  one  considers  the  totality  of  possible  realizations  from  one 
algorithms  vs  another  then  the  two  are  equivalent  if  the  two  sets  are  the  same. 
Unfortunately  the  algorithms  in  general  only  depend  on  second  moment  properties 
and  hence  do  not  uniquely  determine  a  random  function.  One  might  also  consider 
equivalence  in  a  numerical  sense,  two  algorithms  are  equivalent  if  for  any  finite  set 
of  grid  points,  the  same  sets  of  values  are  generated  by  both  algorithms.  While 
equivalence  is  principally  a  theoretical  question,  if  two  algorithms  are  not  equivalent 
then  one  must  ask  how  to  decide  between  algorithms  and  to  what  extent  is  the 
validity  of  the  conclusions  related  to  the  choice  of  the  algorithm. 

All  of  the  simulation  techniques  depend  in  a  critical  way  on  the  use  of  a  random 
number  generators.  It  is  well-known  that  such  numbers  are  in  fact  not  really  random 
but  by  knowing  the  "seed"  the  sequence  is  completely  predictable.  However  there 
are  more  serious  problems,  there  are  many  different  random  generator  algorithms 
and  they  are  not  equivalent  nor  do  they  have  the  same  characteristics.  Some  are 
highly  dependent  on  hardware  and  all  depend  on  the  word  length  in  memory.  A 
highly  touted  "portable"  random  number  generator  has  recently  been  shown  not  to 
be  machine  independent  as  earlier  believed.  As  yet  there  has  been  little  study  of  the 
effect  of  different  random  number  generators  on  the  various  simulation  algorithms. 


28 


REFERENCES 


Alabert,  F.  1987.  The  practice  of  fast  conditional  simulations  through  the  LU 
decomposition  of  the  covariance  matrix.  Math.  Geology  19,  369-386 

Bierkens,  M.A.P.,  and  Weerts,  H.J.T.,  1994,  Application  of  indicator  simulation  to 
modelling  the  lithological  properties  of  a  complex  confining  layer.  Geoderma  62, 
265-284 

Booker,  P.I.  1985.  Two  dimensional  simulation  by  Turning  Bands.  Math.  Geology 
17,81-90 

Dalle-Gupta,  A.,  Lake,  L.  and  Pope,  G.A.,  1995,  Characterizing  heterogeneous 
permeable  media  with  spatial  statistics  and  tracer  data  using  sequential  simulated 
annealing.  Math.  Geology  27,  763-788 

Davis,  M.W.  1987.  Production  of  conditional  simulations  via  the  LU  triangular 
decomposition  of  the  covariance  matrix.  Math.  Geology  19,  91-98 
Davis,  M.W.  1987.  Generating  large  stochastic  simulations-The  matrix  polynomial 
approximation  approach.  Math.  Geology  19,  99-108 

Dietrich,  C.R.  and  Newsam,  G.N.,  1995,  Efficient  generation  of  conditional 
simulations  by  Chebychev  matrix  polynomial  approximations  to  the  symmetric 
square  root  of  the  covariance  matrix.  Math.  Geology  27,  207-228 
Dimitrakopoulos,  R.  1990.  Conditional  simulation  of  intrinsic  random  functions  of 
order  k.  Math.  Geology  361-380 

Deutsch,  C.V.,  1995,  Introduction  to  the  application  of  simulated  annealing  for 

geostatistical  applications.  GEOSTATISTICS  Newsletter  7,  8-14 

Deutsch,  C.V.  and  Cockerham,  P.W.,  1994,  Practical  considerations  in  the 

application  of  simulated  annealing  to  stochastic  simulation.  Math.  Geology  26,  67- 

82 

Easley,  D.  H.,  Borgman,  L.E.  and  Weber,  D.,  1991,  Monitoring  well  placement 
using  conditional  simulation  of  hydraulic  head.  Math.  Geology  23,  1059-1080 
Gutjahr,  A.,  Bullard,  B.  and  Hatch,  S.,  Joint  conditional  simulations  and  flow 
modeling,  in  Geostatistics  for  the  Next  Century,  R.  Dimitrakopoulos  (ed),  Kluwer 
Academic  Pub.  Dordrecht  185-196 

Journel,  A.G.  1974  Geostatistics  for  conditional  simulation  of  ore  bodies.  Economic 
Geology  69 

Journel,  A.G.  and  Isaaks,  E.H.  1984.  Conditional  indicator  simulation:  Application 

to  a  Saskatchewan  uranium  deposit.  Math  Geology  16,  685-718 

Mantoglou,  A.  and  Wilson,  J.L.  1982.  The  turning  bands  method  for  simulation  of 

random  fields  using  line  generation  by  a  spectral  method.  Water  Resources  Res.  1 8, 

1379-1384 

Myers,  D.E.,  1989,  Vector  conditional  simulation,  in  Geostatistics,  M.  Armstrong 
(ed),  Kluwer  Academic  Publishers,  Dordrecht,  283-292 


29 


The  Components  of 
Geostatistical  Simulation 

Carol  A.  Gotway1  and  Brian  M.  Rutherford2 

Abstract.  There  are  many  approaches  to  geostatistical  simulation  that  can 
be  used  to  generate  realizations  of  random  fields.  These  approaches  differ 
fundamentally  in  a  number  of  ways.  First,  each  approach  is  inherently 
different  and  will  produce  fields  with  different  statistical  and  geostatistical 
properties.  Second,  the  approaches  differ  with  respect  to  the  choice  of  the 
features  of  the  region  that  are  to  be  modeled,  and  how  closely  the  generated 
realizations  reproduce  these  features.  Some  fluctuation  in  the  statistical  and 
geostatistical  properties  of  different  realizations  of  the  same  random  field 
is  natural  and  desirable,  but  the  proper  amount  of  deviation  is  an  open 
question.  Finally  the  approaches  differ  in  how  the  conditioning  information 
is  incorporated.  Depending  on  the  source  of  randomness  and  the 
uncertainty  in  the  given  data,  direct  conditioning  of  realizations  is  not 
always  warranted.  In  this  paper,  we  discuss  and  illustrate  these  differences 
in  order  to  emphasize  the  importance  of  these  components  in  geostatistical 
simulation. 


INTRODUCTION 


Geostatistical  simulation,  as  illustrated  in  Figure  1,  provides  a  way  to  quantify 
uncertainty  in  the  prediction  of  a  complex  system  response.  Some  information  is 
usually  available  on  a  process  of  interest  (e.g.,  measurements  on  a  hydrogeologic 
parameter  of  interest  in  a  particular  region),  but  the  transfer  function  (e.g.,  a 
groundwater  flow  model)  may  require  a  detailed  spatial  map  of  the  parameter  on 
which  the  measurements  were  obtained.  (Such  a  map  can  be  created  through  spatial 
prediction  techniques  such  as  kriging.)  The  value  of  the  transfer  function  can  then 
be  computed  over  the  mapped  values  to  obtain  a  prediction  of  a  system  response 
(e.g.,  groundwater  travel  time),  but  an  estimate  of  the  uncertainty  associated  with  the 
prediction,  often  as  important  as  the  predicted  value  itself,  is  typically  much  more 
difficult  to  infer.  One  alternative  is  to  generate  realizations  of  the  random  field  that 
share  the  available  information  on  the  parameter  of  interest  and  use  the  transfer 
function  to  compute  a  system  response  for  each.  If  the  realizations  characterize  the 
spatial  uncertainty  in  the  parameters  of  interest,  the  resulting  distribution  of  predicted 
system  responses  will  reflect  this  uncertainty. 

1  Assistant  Professor  of  Biometry,  University  of  Nebraska,  Lincoln,  NE 

2  Senior  Member  of  Technical  Staff,  Sandia  National  Laboratories,  Albuquerque,  NM 


30 


Multiple  Transfer  Distribution  of 

Realizations  Function  System  Response 


Figure  1.  Illustration  of  Stochastic  Simulation  (Journel,  1989). 


Although  originally  introduced  as  a  geostatistical  approach  to  uncertainty  analysis 
(Journel,  1989),  geostatistical  simulation  has  grown  to  provide  a  variety  of  new 
approaches  to  stochastic  modeling  that  can  incorporate  complex  spatial  relationships. 
These  methods  are  frequently  used  in  hydrology,  petroleum  engineering,  and  the 
environmental  sciences,  not  only  for  uncertainty  analysis,  but  also  for  risk  analysis, 
decision  analysis,  and  stochastic  modeling. 

Many  different  approaches  to  geostatistical  simulation  can  be  used  to  construct  the 
realizations.  For  continuous  variables  these  include  the  turning  bands  method 
(Journel,  1974),  spectral  methods  (Borgman  et  al.,  1984;  Gutjahr,  1989),  LU  or 
Cholesky  decomposition  (Alabert,  1987;  Davis,  1987),  sequential  Gaussian 
simulation  (Journel,  1989),  and  fractal  approaches  (Hewett  and  Behrens,  1990). 
Approaches  to  simulation  of  discrete  or  categorical  variables  include  Boolean 
methods  (Serra,  1982;  Jeulin,  1987;  Chautru,  1989)  and  the  more  general  random 
closed  set  approaches  (Stoyen  et  al.,  1987),  and  the  use  of  truncated  Gaussian 
random  functions  (Matheron  et  al.,  1987).  Sequential  indicator  simulation  (Journel 
and  Alabert,  1989),  simulated  annealing  (Kirkpatrick  et  al.,  1983;  Deutsch  and 
Journel,  1992)  and  genetic  algorithms  (Whitley,  1994)  can  be  used  with  both 
continuous  or  categorical  variables. 

For  any  particular  type  of  data  (continuous,  discrete,  or  categorical),  different 
approaches  will  assign  different  statistical  and  spatial  features  to  the  generated  fields. 
These  differences  could  be  due  to:  (1)  differences  in  the  statistical  model  used  to 
characterize  the  phenomenon  under  study.  These  differences  are  most  obviously 
represented  by  the  criteria  used  to  describe  spatial  data  relationships  and  characterize 


31 


available  information.  We  refer  to  such  criteria  as  summary  functions.  Examples  of 
summary  functions  include  the  univariate  data  distribution,  the  covariance  function, 
indicator  semivariograms,  and  the  shape  and  intensity  of  primary  elements  in 
Boolean  simulations;  (2)  differences  in  algorithms  (e.g.,  using  sequential  Gaussian 
simulation  versus  using  LU  decomposition);  (3)  the  acceptable  deviation  between 
the  summary  functions  specified  to  the  algorithm  and  those  calculated  from  resulting 
realizations;  and  (4)  the  use  of  conditioning  information.  Since  all  of  these 
differences  will  translate  into  differences  in  the  nature  of  the  spatial  heterogeneity 
reflected  in  the  resulting  realizations,  and  hence  into  differences  in  uncertainty 
assessments,  risk  calculations,  and  the  conclusions  drawn  from  them,  it  is  important 
to  consider  each  of  these  components  carefully  when  using  geostatistical  simulation. 
In  the  remainder  of  this  paper,  we  elaborate  on  these  components,  discuss  some  of 
the  different  choices  within  each  component,  and  illustrate  the  effect  of  these 
differences  on  the  nature  of  the  simulated  fields  and  on  functions  derived  from  them. 


COMPONENTS  OF  GEOSTATISTICAL  SIMULATION 


When  using  geostatistical  simulation  to  characterize  the  spatial  variation  of 
parameters  of  interest,  the  modeler  has  to  make  several  important  choices.  Each 
choice  will  affect  the  nature  of  the  resulting  realizations  and  the  values  of  functions 
derived  from  them.  Thus,  careful  thought  with  regard  to  the  major  components  of 
geostatistical  simulation  is  crucial  in  modeling  the  uncertainty  of  spatial  functions. 

Choice  of  statistical  model  and  selection  of  summary  functions 

For  geostatistical  simulation  of  a  spatial  process,  a  statistical  model  must  describe 
a  multivariate  probability  distribution  for  the  spatial  data  of  interest.  Given  such  a 
distribution,  realizations  can  be  generated  by  sampling  randomly  from  this  specified 
distribution.  For  example,  in  LU  decomposition,  the  underlying  multivariate 
distribution  is  Gaussian,  and  the  parametric  form  for  this  distribution  can  be  found 
in  most  introductory  statistical  texts.  Similarly,  general  formula  for  the  multivariate 
distributions  of  Boolean  random  sets,  as  well  as  formula  for  distributions  of  other 
random  sets,  are  given  in  Stoyen  et  al.,  (1987).  In  other  cases,  such  as  the  sequential 
indicator  simulation  model,  or  certain  Markov  Random  Field  Models,  the  exact  form 
of  the  multivariate  distribution  is  intractable  or  unknown,  and  is  conditionally 
specified  using  a  local  neighborhood  structure. 

Once  a  particular  model  has  been  chosen,  the  multivariate  distribution  is  specified 
through  the  use  of  summary  functions.  For  example,  generation  of  realizations  from 
a  multivariate  Gaussian  distribution  requires  specification  of  the  process  mean  and 


32 


covariance  structure  which  is  typically  determined  through  the  semivariogram.  In 
Boolean  models,  the  multivariate  distribution  depends  on  the  intensity  of  the  Poisson 
process,  the  shape  of  the  primary  elements,  and  the  probability  distributions  of  their 
sizes.  In  sequential  indicator  simulation,  the  multivariate  distribution  is  specified 
nonparametrically  through  the  use  of  indicator  semivariograms.  Consequently,  using 
a  particular  simulation  algorithm  that  requires  a  specified  set  of  summary  functions 
is  equivalent  to  specifying  a  multivariate  probability  model  for  the  spatial  process. 
These  models  necessarily  emphasize  different  spatial  features  of  the  process,  so  it  is 
important  to  determine  what  features  are  most  important  in  the  application  of  interest 
and  how  best  to  quantify  them  from  the  available  information.  In  applications  where 
a  transfer  function  is  evaluated  over  the  generated  realizations,  it  is  important  to 
determine  the  features  of  the  process  that  will  greatly  influence  the  value  calculated 
through  the  transfer  function.  For  example,  in  reservoir  modeling,  it  is  often 
important  to  characterize  those  features  that  influence  the  flow  of  fluids  through 
reservoirs,  and  in  environmental  risk  analysis,  it  may  be  important  to  characterize 
those  features  that  control  the  spread  of  a  contaminant  of  special  interest. 

Emphasizing  different  features  through  the  use  of  different  summary  functions  can 
give  rise  to  realizations  that  appear  quite  different.  As  an  example,  consider  the  top 
image  in  Figure  2.  It  shows  a  geologic  section  (obtained  from  digitization)  of 
interbedded  debris  flow  (fine  grained)  and  sheet  flood  (coarse  grained)  units 
produced  as  part  of  site-characterization  studies  at  the  Greater  Confinement  Disposal 
Facility  in  Southern  Nevada  (Rutherford  and  Gotway,  1994;  McCord  et  al.,  1996). 
Recharge  through  sections  of  this  type  is  an  important  component  of  such  studies  and 
the  position,  shape,  and  amount  of  each  type  of  unit  are  important  for  modeling 
vadose  zone  flow.  The  last  three  images  in  Figure  2  are  unconditional  realizations 
generated  from  statistics  inferred  from  the  training  image  using  sequential  indicator 
simulation.  Boolean  simulation,  and  a  more  general  random  set  approach.  Sequential 
indicator  simulation  uses  two  summary  functions,  the  proportion  of  debris  flow  and 
the  indicator  semivariogram.  The  indicator  semivariogram  attempts  to  incorporate 
the  connectivity  of  the  permeable  and  impermeable  zones.  Although  the  shape  of  the 
debris  flow  units  can  be  controlled  to  some  degree  by  the  anisotropy  reflected  in  the 
indicator  semivariograms,  the  shapes  are  somewhat  limited.  The  Boolean  model 
uses  summary  functions  that  specify  the  intensity  of  the  process,  and  the  shape  and 
size  of  the  debris  flow  units.  Here  the  emphasis  is  on  shape,  rather  than  connectivin . 
The  last  realization  is  based  on  the  theory  of  random  closed  sets  (Kendall,  1 974).  In 
this  application  the  summary  functions  are  described  by  a  system  of  structuring 
elements,  and  any  random  set  can  be  characterized  through  its  intersection  with  these 
elements.  The  last  realization  was  generated  using  19  structuring  elements 
consisting  of  lines,  triangles  and  rectangles.  Here  there  is  an  attempt  to  account  for 
both  shape  and  connectivity.  The  use  of  multiple-point  statistics  (Deutsch  and 
Journel.  1992;  Guardiano  and  Srivastava.  1993)  might  also  be  used  to  incorporate 


33 


GCD  Training  Image 


Sequential  Indicator  Realization 


Boolean  Realization 


Hitting  Functions  Realization 


Figure  2.  The  Use  of  Different  Summary  Functions  in  Lithologic  Simulation. 


34 


more  features  of  importance.  Note  that  in  such  applications,  the  goal  is  not  to 
reproduce  the  training  image,  but  to  capture  essential  features  of  the  image  (not 
always  visible  by  eye)  that  are  most  important  to  the  transfer  function  of  interest. 


Choice  of  Simulation  Algorithm 


In  addition  to  the  selection  of  an  appropriate  set  of  summary  functions  that 
determine  the  modeling  approach,  different  algorithms  for  generating  the  realizations 
often  impact  the  results.  In  implementing  most  geostatistical  simulation  algorithms, 
there  are  a  number  of  parameters  that  must  be  made  specific,  often  based  on  a 
tradeoff  betw  een  speed  and  accuracy.  Examples  include  the  number  of  turning  bands 
lines,  the  number  of  conditioning  values  retained  in  sequential  simulation,  the  choice 
of  annealing  schedule  and  convergence  criteria,  and  the  number  of  structural 
elements  in  random  set  generation.  In  many  cases,  blatantly  improper  choices  can 
be  clearly  reflected  in  the  realizations,  e.g.,  the  appearance  of  banding  effects  in 
turning  bands  method  when  the  number  of  lines  is  small.  In  other  cases,  apparently 
reasonable  selections  can  also  have  an  effect  on  the  realizations  and  on  the  resulting 
uncertainty  distributions  derived  from  them.  In  a  comparative  study  of  several 
geostatistical  simulation  algorithms  using  a  variety  of  exhaustive  data  sets  and 
transfer  functions,  Gotway  and  Rutherford  (1994)  illustrate  these  algorithmic 
differences.  Figure  3  shows  uncertainty  distributions  of  the  average  of  local 
geometric  means  computed  from  conditional  realizations  of  a  Gaussian  field. 
Although  the  shape  of  the  uncertainty  distributions  is  similar,  they  differ  in  the  upper 
tail  which  could  be  crucial  for  risk  assessments.  Comparing  the  distributions 
obtained  using  LU  Decomposition,  Sequential  Gaussian,  and  Turning  Bands 
algorithms  shows  the  effect  of  algorithmic  differences  on  the  response  uncertainty 
distributions.  Comparing  the  distribution  obtained  using  sequential  indicator 
simulation  to  the  others  illustrates  the  effect  of  choice  of  summary  functions.  As 
shown  in  Gotway  and  Rutherford  (1994),  these  differences  are  much  more 
pronounced  for  unconditional  realizations. 


The  Acceptable  Deviation 


The  acceptable  deviation  refers  to  the  difference  between  the  summary  functions 
calculated  from  realizations  and  those  specified  by  the  model.  For  example,  should 
the  variogram  calculated  from  each  realization  exactly  match  the  variogram 
specified?  Certainly,  the  answer  to  this  question  is  no.  since,  in  simulation,  some 
deviation  of  summary  functions  is  desirable  in  order  to  reflect  uncertainty.  However, 
the  amount  of  deviation  acceptable  is  an  open  question.  Unless  there  is 


35 


Percent 
40r 


30 


20 


1.0 


LU  Decomposition 


5.0 

Response  Value 


10.0 


Percent 
40r 


30 


20 


1.0 


Sequential  Gaussian 


5.0 

Response  Value 


10.0 


Percent 
40r 


30 


20 
10 


1.0 


Sequential  Indicator 


5.0 

Response  Value 


10.0 


Percent 
40r 


30 


20 


1.0 


Turning  Bands 


a 


5.0 

Response  Value 


10.0 


Figure  3.  Response  Uncertainty  Distributions. 

"post-generation"  alteration  to  the  realizations,  these  deviations  depend  on  the 
specified  summary  functions  and  on  the  simulation  algorithm.  Often,  simulation 
algorithms  are  compared  and  evaluated  based  on  their  ability  to  reproduce  the  desired 
functions  over  an  average  of  realizations.  However,  the  variability  in  the  summary 
functions  can  be  quite  different  depending  on  the  particular  algorithm  used  and  still, 
on  the  average,  reproduce  the  specified  summary  functions.  More  importantly,  such 
a  reproduction  assumes  that  the  summary  functions  specified  are  in  fact  known, 
when  in  reality  they  are  estimated  from  available  data. 

The  fundamental  concern  with  the  acceptable  deviation  is  that  differences  between 
specified  summary  functions  and  those  calculated  from  generated  realizations  should 
reflect:  (1)  the  uncertainty  in  input  summary  functions;  (2)  the  natural  variability 
that  comes  with  random  spatial  sampling  over  a  finite  region  to  generate  realizations, 
and  not  (3)  the  variability  (or  lack  of  variability)  introduced  by  the  simulation 
algorithm.  None  of  the  available  algorithms  account  for  (1),  and  with  the  exception 
of  algorithms  in  which  a  multivariate  distribution  is  directly  sampled  (such  as  LU 
decomposition)  it  is  also  not  clear  whether  (2)  is  accomplished  correctly.  For  many 
applications,  (3)  may  be  the  primary  determinate  of  deviations  between  theoretical 
and  reproduced  summary  functions. 


36 


The  Use  of  Conditioning  Information 


Conditioning  a  simulation  on  known  values  or  soft  information  (incomplete 
information  concerning  the  region  of  interest)  must  be  done  in  a  way  that  reflects  the 
true  conditions  imposed  on  the  simulated  region  by  that  data  or  information.  At 
present,  the  common  approach  for  algorithms  using  a  sequential  method  of 
generation  is  to  treat  previously  generated  values  as  known  data  values  in  the 
simulation.  The  study  in  Hansen  (1992)  indicates  that  this  approach  might  "over- 
condition"  or  "under-condition"  the  realizations.  The  result  is  a  biased  uncertainty 
distribution  of  response  that  may  be  unrealistically  wide  or  too  narrow.  Moreover, 
if  data  are  measured  with  error,  conditioning  on  these  data  may  not  be  warranted; 
rather  some  filter  must  first  be  applied  to  reduce  the  noise  introduced  by 
measurement  error. 

In  many  applications,  the  summary  functions  are  derived  from  a  training  image, 
and  simulation  of  complex  geologic  structures  is  often  based  on  a  training  image  that 
is  derived  from  outcrop  mapping  (e.g.,  Figure  2).  However,  the  goal  of  simulation 
is  not  to  reproduce  a  training  image.  Rather,  the  objective  is  to  generate  realizations 
that  account  for  information  believed  relevant  to  the  general  process  under  study. 
Such  information  might  include  the  general  shape  and  relative  distributions  of 
important  units,  connectivity  patterns,  cyclic  trends,  etc.,  that  could  be  expected  to 
be  general  properties  of  the  process.  It  most  probably  does  not  include  exact  values 
obtained  from  the  training  image.  In  fact,  Guardiano  and  Srivastava  (1 993)  note  that 
"the  training  image  is  itself  is  not  the  underlying  phenomenon.  It  is  assumed, 
however,  that  both  share  some  structural  properties  such  as  specific  multiple-point 
covariances."  Consequently,  it  seems  very  difficult  to  justify  using  the  exact  values 
(or  even  some  randomly  selected  subset  of  them)  as  conditioning  information. 

REFERENCES 

Alabert,  F.G.  1987.  The  practice  of  fast  conditional  simulations  through  the  LU 
decomposition  of  the  covariance  matrix.  Mathematical  Geology,  19:  369-386. 

Borgman,  L.E.,  M.  Taheri,  and  R.  Hagen.  1984.  Three-dimensional,  frequency- 
domain  simulations  of  geological  variables.  In  G.  Verly  et  al.  (eds.), 
Geostatistics  for  Natural  Resources  Characterization.  Reidel.  Dordrecht, 
pp.  517-541. 

Chautru,  J.M.  1989.  The  use  of  Boolean  random  functions  in  geostatistics.  In 
M.  Armstrong  (ed.),  Geostatistics,  Vol.  1,  Kluwer.  Dordrecht,  pp.  201-212. 

Davis,  M.W.  1987.  Generating  large  stochastic  simulations  via  the  LU  triangular 
decomposition  of  the  covariance  matrix.  Mathematical  Geology.  19:  91-98. 


37 


Deutsch,  C.V.  and  A.G.  Journel.  1992.  Annealing  techniques  applied  to  the 

integration  of  geological  and  engineering  data.  Stanford  Center  for  Reservoir 

Forecasting,  Stanford,  CA. 
Gotway,  C.A.,  and  B.M.  Rutherford.  1994.  Stochastic  simulation  for  imaging 

spatial  uncertainty:  Comparison  and  evaluation  of  available  algorithms.  In 

M.  Armstrong  and  P.A.  Dowd  (eds.),  Geostatistical  Simulations,  Kluwer, 

Dordrecht,  pp.  1-21. 
Guardiano,  F.  And  M.  Srivastava.  1993.  Multivariate  geostatistics:  Beyond 

bivariate  moments.  In  Soares,  A.  (ed.),  Geostatistics  Troia  '92.  Kluwer, 

Dordrecht,  pp.  133-144. 
Gutjahr,  A.  1989.  Fast  Fourier  Transforms  for  Random  Fields.  Technical  Report 

No.  4-R58-2690R,  Los  Alamos,  NM. 

Hansen,  K.M.  1992.  The  use  of  sequential  indicator  simulation  to  characterize 

geostatistical  uncertainty.  Technical  Report  No.  SAND91-0758,  Sandia 

National  Laboratories,  Albuquerque,  NM. 
Hewett,  T.  And  R.  Behrens.  1990.  Conditional  simulation  of  reservoir 

heterogeneity  with  fractals.  Formation  Evaluation,  330-310. 
Jeulin,  D.  1987.  Anisotropic  rough  surface  modeling  by  random  morphological 

function.  Acta  Stereologica,  6: 1 83-1 89. 
Journel,  A.G.  1974.  Geostatistics  for  conditional  simulation  of  ore  bodies. 

Economic  Geology,  69:673-687. 
Journel,  A.G.  1989.  Fundamentals  of  Geostatistics  in  Five  Lessons.  American 

Geophysical  Union,  Washington,  D.C. 
Journel,  A.G.,  and  Alabert,  F.  (1989).  Non-Gaussian  data  expansion  in  the  earth 

sciences.  Terra  Nova,  1:123-124. 
Kendall,  D.G.  1974.  Foundations  of  a  theory  of  random  sets.  In  E.  Harding  and 

D.  Kendall  (eds.),  Stochastic  Geometry,  Wiley,  Chichester,  pp.  322-376. 
Kirkpatrick,  S.,  CD.,  Gelatt,  and  M.P.  Vecchi.  1983.  Optimization  by  simulated 

annealing.  Science,  220:67 '1  -680. 
Matheron,  G.,  H.  Beucher,  C.  De  Fouquet,  A.  Galli,  D.  Guerillot,  and  C.  Ravenne. 

1987.  Conditional  simulation  of  the  geometry  of  fluvio-deltaic  reservoirs.  SPE 

Paper  No.  16753,  62nd  Annual  Technical  Conference  and  Exhibition  of  the 

Society  of  Petroleum  Engineers,  Dallas,  TX.,  September  27-30. 
McCord,  J.T.,  CA.  Gotway,  and  S.H.  Conrad.  1996.  Inferring  recharge  rates 

using  environmental  tracers  in  heterogeneous  vadose  zones.  Water  Resources 

Research,  to  appear. 
Rutherford,  B.M.,  and  CA.  Gotway.  1994.  Selecting  features  from  spatial  data 

for  use  in  stochastic  simulation.  Proceedings  of  the  35th  Annual  Meeting  of  the 

Institute  of  Nuclear  Materials  Management,  July  17-20,  Naples,  FL. 
Serra,  J.  1982.  Image  Analysis  and  Mathematical  Morphology.  Academic 

Press,  London. 

Stoyen,  D.,  W.S.  Kendall,  and  J.  Mecke.  1987.  Stochastic  Geometry  and  Its 
Applications.  Wiley,  New  York. 

Whitley,  D.  1994.  A  genetic  algorithm  tutorial.  Statistics  and  Computing, 
4:  65-85. 

38 


A  General  Method  for  the 
Efficient  Selection  of  Sampling 
Locations  for  Problems  in 
Environmental  Restoration 

Brian  M.  Rutherford1 


Abstract -Problems  in  environmental  restoration  that  involve  detecting  or 
monitoring  contamination  or  site  characterization  often  benefit  from 
procedures  that  help  select  sampling  or  drilling  locations  for  obtaining 
meaningful  data  that  support  the  analysis.  One  example  of  this  type  of 
procedure  is  a  spatial  sampling  program  that  will  "automatically"  (based 
on  the  implementation  of  a  computer  algorithm)  guide  an  iterative 
investigation  through  the  process  of  site  characterization  at  a  rninimal  cost 
to  determine  appropriate  remediation  activities.  In  order  to  be  effective, 
such  a  procedure  should  translate  site  and  modeling  uncertainties  into 
terms  that  facilitate  comparison  with  regulations  and  should  also  provide  a 
methodology  that  will  lead  to  an  efficient  sampling  plan  over  the  course  of 
the  analysis.  In  this  paper,  a  general  framework  is  given  that  can 
accomplish  these  objectives  and  can  be  applied  to  a  wide  range  of 
environmental  restoration  applications.  The  methodology  is  illustrated 
using  an  example  where  soil  samples  support  the  characterization  of  a 
chemical  waste  landfill  area. 


INTRODUCTION 

The  objectives  of  environmental  restoration  projects  often  require  data  for 
detecting,  monitoring,  or  estimating  the  extent  of  contamination.  Spatial  sampling 
procedures  may  be  required  to  help  locate  soil  sample  or  bore  hole  locations  that 
will  provide  the  most  additional  site  information  at  the  lowest  cost.  Two  examples 
of  problems  that  might  benefit  from  this  type  of  procedure  follow: 

a)  Find  drilling  locations  for  a  monitoring  system  that  will  provide  the 
information  required  to  minimize  the  probability  that  a  contaminant  will  cross  a  site 
boundary  undetected;  and 

b)  Select  samples  to  reduce  as  much  as  possible  the  area  or  volume  of  a  90% 
confidence  region  for  a  specific  concentration  contour. 


1  Senior  Member  Technical  Staff,  Sandia  National  Laboratories,  Albuquerque,  NM. 


39 


Two  essential  elements  of  procedures  designed  to  accomplish  these  types  of 
environmental  restoration  project  objectives  are:  I)  to  translate  site  and  modeling 
uncertainties  into  terms  that  permit  comparison  with  regulations  so  that 
compliance  and  cleanup  requirements  can  be  assessed;  and  II)  to  provide 
methodology  that  can  lead  to  an  efficient  sample  plan  throughout  the  course  of  the 
analyses.  The  concepts  involved  with  these  two  elements  of  the  procedure  provide 
the  basis  for  the  remainder  of  this  paper. 

TRANSLATING  SITE  AND  MODELING  UNCERTAINTIES 
INTO  TERMS  OF  REGULATORY  COMPLIANCE 


Limited  sampling  and  other  factors  contribute  to  site  and  modeling  uncertainty. 
Most  site  characterization  applications  include  uncertainty  from  a  number  of 
sources:  contaminant  values  at  locations  that  have  yet  to  be  sampled  are  unknown; 
flow  and  geostatistical  models  have  uncertainty  associated  with  them;  and  data 
available  from  samples  collected  at  the  site  and  "soft  data"  relevant  to  the  site, 
collected  through  a  number  of  alternative  characterization  methodologies  have 
varying  degrees  of  uncertainty  associated  with  them. 


Site  Knowledge 


Different  conceptual  models 


Probability  of  performing 
remediation  activities 


Uncertainties  in  parameter  and 
state  variable  values  at 
un  sampled  locations 


Conditioning  data 


Compliance 
Requirements 


^  1 

True  Average  Concentration 


Figure  1 

Recognizing  these  uncertainties  as  unavoidable,  environmental  regulations 
should  be  and  often  are  stated  probabilistically.  Constraints  on  the  allowable 
probability  of  exceeding  a  regulatory  threshold,  stated  as  a  function  of  the 
(hypothetical  unknown)  true  level  of  contamination,  provides  one  example.  In 
order  to  be  effective,  the  procedure  must  translate  site  uncertainties  into  the 
probabilistic  framework  of  these  regulations.  Figure  1  illustrates  this  element  of 
the  process  for  the  (simpler  than  most)  problem  of  characterizing  average 
contamination  levels.  This  will  provide  those  responsible  for  demonstrating 
compliance  of  a  site,  or  those  responsible  for  analyzing  cleanup  alternatives  and 
capabilities,  with  a  way  of  making  critical  decisions. 


40 


The  examples  discussed  earlier  will  be  different  (and  generally  more  complex). 
Example  a),  for  example,  might  have  a  vertical  scale  indicating  the  detection 
probability.  The  computation  of  these  probabilities  would  depend  on  the  assumed 
flow  models  and  their  parameters  as  well  as  the  geological  features  of  the  site. 

This  translation  of  site  uncertainties  into  a  regulatory  framework  can  be 
achieved  through  a  two-step  process.  The  first  step  is  to  convert  site  and 
modeling  uncertainties  into  simulated  realizations,  each  realization  providing  a 
possible  map  of  site  contamination.  When  taken  collectively,  the  realizations  reflect 
probabilistically  the  way  that  contaminant  levels  at  the  site  might  be  configured. 
These  realizations  are  then  evaluated  according  to  an  objective  function  that 
calculates  the  compliance  probability  or  risk. 

The  remainder  of  this  section  is  partitioned  into  subsections  that  address  these 
three  components  of  this  process 

Factors  Contributing  to  Site  and  Modeling  Uncertainty 

At  least  five  general  sources  might  contribute  to  uncertainty  in  the  estimated 
concentration  levels  of  contaminants  at  a  site: 

1)  Uncertainties  in  small  scale  hydrogeologic  property  values,  either  in 
parameter  values  or  in  state  variables.  These  values  are  uncertain  at  locations  that 
have  not  been  sampled; 

2)  Uncertainty  introduced  by  incomplete  knowledge  regarding  the  location 
and  extent  of  large  scale  geological  features  such  as  stratigraphic  units  or  zones  of 
heavy  fracturing; 

3)  Uncertainty  concerning  the  extent  and  configuration  of  the  contaminant 
source; 

4)  Uncertainty  concerning  the  appropriate  flow  and  transport  models  and  in 
parameters  associated  with  these  models;  and 

5)  Uncertainty  in  analytical  results  for  samples  already  collected  and  analyzed. 

The  first  step  of  the  translation  process  is  to  convert  site  and  modeling 
uncertainties  into  simulated  realizations  that  differ  from  one  another  to  an  extent 
that  reflects  these  uncertainties.    Each  source  listed  above  must  be  taken  into 


41 


account  in  the  simulation  process  to  reflect  its  impact  on  differences  between 
simulated  realizations. 

Fields  of  hydrogeologic  parameter  values  can  be  generated  using  stochastic 
simulation  techniques  conditioned  on  sample  data  where  available.  Replication  of 
this  simulation  process  can  account  for  parameter  uncertainty  source  (1). 
Uncertainties  in  state  variables  are  often  estimated  through  repeated  application  of 
the  flow  models  using  different  sets  of  the  hydrogeological  and  other  parameter 
values  (selected  through  simulation)  for  each  replication  and  conditioned  based  on 
known  values  of  state  variables  where  applicable.  Other  parameter  values  may 
include  the  specification  of  large  scale  geological  features,  source  (2)  above,  or 
contaminant  source,  strength  and  configuration,  source  (3).  Probabilities  for 
different  contaminant  sources  and  for  different  scenarios  involving  large  scale 
geological  features  must  be  assessed  (perhaps  subjectively)  and  introduced  into  the 
simulation  process  accordingly. 

Modeling  uncertainty,  source  (4),  can  be  accounted  for  by  selecting  among 
alternative  models  with  probability  of  selection  consistent  with  each  model's 
likelihood  of  being  applicable  (again,  likely  to  be  a  subjective  decision).  In  most 
cases,  the  appropriate  models  are  established,  but  parameters  associated  with  these 
models  must  be  specified.  Their  uncertainty  distribution  must  be  estimated,  and 
the  parameters  generated  randomly  through  Monte  Carlo  or  perhaps  alternative 
simulation  techniques.  The  fifth  source  of  uncertainty  —  that  resulting  from  inexact 
analytical  procedures,  can  be  accounted  for  by  reducing  the  conditioning 
requirements  of  the  simulation  and  consequently  allowing  differences  between 
values  obtained  through  sampling  and  those  obtained  through  simulation  at  the 
same  location. 

Together,  accommodating  for  these  sources  of  uncertainty  through  the 
simulation  process  will,  if  performed  correctly,  provide  an  estimated  probability 
space  for  site  reality  that  is  both  unbiased  and  provides  an  accurate  assessment  of 
uncertainty  based  on  present  site  knowledge.  This  discrete  (by  necessity) 
representation  of  the  probability  space  is  described  next. 

The  Simulated  Realizations 

The  simulated  realizations  represent  an  intermediate  step  in  the  translation 
process.  Each  realization  provides  one  possible  map  of  site  contaminant  levels.  At 
each  grid  location,  the  set  of  realizations  yields  a  histogram  of  possible  values. 
Figure  2  illustrates  how  the  set  of  realizations  can  be  interpreted  in  this  way. 


42 


Multiple 
Realizations 


Figure  2 

An  appropriate  interpretation  of  this  set  of  realizations,  analyzed  pointwise,  is 
that  the  probability  of  a  contaminant  value  at  this  location,  smaller  than  any  level 
x,  can  be  estimated  by  the  ratio  of  the  number  of  realizations  that  have  a  value  less 
than  x  at  this  location,  to  the  total  number  of  realizations.  This  probabilistic 
interpretation  can  be  extended  to  multiple  locations  or  to  the  entire  region.  The 
probability  that  no  point  in  the  entire  region  exceeds  the  level  y,  for  example,  can 
be  estimated  by  the  ratio  of  the  number  of  realizations  where  y  is  not  exceeded  at 
any  point  in  the  region  to  the  total  number  of  realizations. 

This  interpretation  of  the  set  of  realizations  is  utilized  in  two  ways  in  the 
methodology  presented  in  this  report.  It  provides  justification  for  the  probabilistic 
interpretation  of  the  values  indicting  likelihood  of  compliance  or  risk  functions  that 
are  compared  to  regulations  and  it  provides  a  basis  for  the  resampling  procedure 
discussed  in  the  subsection  Determining  an  Efficient  Sample  Plan. 

The  Objective  Function 

The  objective  function  is  a  transformation  that  will  compute  a  value  indicating 
the  level  of  compliance  based  on  a  complete  specification  of  all  site  parameters  and 
state  variables.  For  the  present  problem,  the  objective  function  is  used  to  transform 
the  set  of  simulated  realizations  into  a  probabilistic  expression  that  relates  to 
regulatory  compliance.  Examples  of  the  roles  that  objective  functions  might  play  in 
the  example  problems  listed  in  the  Introduction  Section  are: 

a)  Provide  an  estimate  of  the  probability  that  a  contaminant  will  cross  a  site 
boundary  undetected  by  a  specific  monitoring  system  within  10  years;  and 

b)  Provide  a  distribution  showing  probabilistically  the  spatial  or  volumetric 
extent  of  the  90%  confidence  region  for  a  specific  concentration  contour. 


43 


DETERMINING  AN    EFFICIENT  SAMPLING  PLAN 


The  process  of  sample  location  selection  is  often  an  iterative  process.  As  new 
data  are  obtained,  models  are  updated  to  provide  more  accurate  estimates  of  site 
contamination  levels  or  predictions  of  flow  and  transport  and  a  reassessment  of 
future  data  needs  is  performed.  A  second  essential  element  of  an  effective  spatial 
sampling  procedure  is  that  it  utilizes  data  available  at  every  sampling  iteration  and 
that  models  based  on  these  data  produce  an  efficient  sample  plan  —  one  where 
maximum  relevant  information  is  provided  at  the  lowest  possible  cost.  At  each 
iteration,  the  next  set  of  sample  locations  is  selected  from  numerous  potential 
supplemental  sample  sets.  Methods  for  choosing  sets  to  consider  and  methods  for 
evaluating  those  that  are  considered  must  be  carefully  selected. 

A  number  of  applications  have  been  reported  in  the  hydrology  and  geology 
literature  where  selection  of  sampling  or  monitoring  locations  has  been  approached 
using  probabilistic  optimization  techniques.  Early  applications,  involving  resource 
assessment  generally  in  mining  applications,  were  based  on  the  assumption  that  the 
level  of  the  state  variables  in  the  region  (ore  grades,  in  many  applications)  could  be 
modeled  as  a  Gaussian  random  field  using  regionalized  variable  theory.  For  these 
applications,  the  utility  of  any  future  sample  set  could  be  assessed  prior  to 
collection  and  analysis  by  estimating  the  reduction  in  kriging  variance  or  block 
kriging  variance  that  the  set  yielded. 

Recently,  environmental  restoration  applications  addressing  spatial  sampling 
problems  have  appeared,  primarily  in  the  hydrology  literature.  These  application 
are  often  less  straight-forward  than  those  based  on  regionalized  variable  theory. 
Most  applications  are  based  on  flow  models  that  use  as  input,  maps  of 
hydrogeological  parameters  (porosity  or  hydraulic  conductivity,  for  examples). 
Another  possible  difference  between  environmental  restoration  problems  and  those 
of  the  earlier  applications  is  the  variety  of  objectives  for  different  analyses.  The 
objective  function  determines  the  utility  of  further  sampling  and  that  of  sampling  at 
specific  sets  of  sample  locations.  The  likelihood  of  a  contaminant  crossing  a  site 
boundary  undetected  by  a  specific  monitoring  system,  for  example,  is  an  analysis 
objective  that  is  not  related  to  a  reduction  in  the  kriging  variance  in  any  obvious 
way.  Consequently,  the  utility  of  additional  sample  or  monitoring  locations  must 
be  approached  differently  for  this  objective. 

A  number  of  the  publications  involving  spatial  sampling  problems  in 
environmental  restoration  focus  on  the  development  of  methods  that  facilitate  the 
assessment  of  state  variable  information  (hydraulic  head  or  contaminant  levels,  for 
example)  provided  by  samples  of  hydrologic  parameters.  Another  primary  focus 
has  been  in  the  development  of  optimization  techniques  that  approach  these 


44 


specific  objectives  in  an  efficient  manner.  One  article  that  provides  an  excellent 
review  of  accomplishments  in  these  areas  and  provides  numerous  further 
references  is  Mc  Kinney  and  Loucks  (1992).  The  primary  objective  in  the  present 
paper  is  to  develop  a  more  general  approach  to  optimization  for  spatial  sampling 
problems  that  can  be  applied  to  a  wide  range  of  problems  regardless  of  their 
modeling  requirements  and  varying  analysis  objectives. 

A  sampling  plan  can  be  constructed  utilizing  the  appropriate  probability  space 
generated  using  the  procedures  outlined  above,  along  with  optimization  methods 
outlined  in  this  section.  A  more  detailed  discussion  of  these  methods  is  provided 
in  Rutherford  (1996).  The  sampling  plan  is  based  on  an  iterative  procedure  that 
requires  evaluation  of  samples  collected  at  each  iteration  of  the  process  to  provide 
new  information  for  selecting  sites  to  be  used  in  the  next  iteration.  Two  issues 
arise  in  this  process  that  will  be  outlined  in  the  next  two  subsections.  The  final 
subsection  provides  an  example  of  the  entire  process. 

Assessing  the  Value  of  Sampling  at  Unsampled  Locations 

As  alternative  sets  of  supplemental  sample  locations  are  considered,  there  is  a 
need  to  evaluate  the  potential  for  improvement  provided  by  each  sample  set  prior 
to  collection  or  analyses  of  the  samples.  This  requires  an  estimation  of  the 
information  available  from  each  sample  set  and  an  evaluation  of  its  likely  impact  on 
the  objectives  of  the  overall  analysis.  Assessment  of  the  relative  overall 
information  provided  by  sampling  at  any  set  of  sample  locations  depends  on  site 
and  modeling  information  that  has  been  translated  to  the  simulated  realizations. 
The  effect  of  "realizing"  values  projected  by  the  simulated  realizations  provides  an 
indication  of  what  might  occur  if  that  sample  set  was  actually  selected  and 
analyzed. 

Figure  2  shows  how  the  resampling  process  might  be  implemented.  For  each 
hypothetical  sample  set,  realizations  (selected  at  random)  are  assumed  to  represent 
the  true  status  of  concentration  at  the  site.  Their  values  at  the  candidate 
supplemental  sample  locations  can  be  submitted  as  conditioning  data  along  with 
known  sample  information  to  formulate  new  models.  These  models  can  be  used  to 
generate  a  new  set  of  simulated  realizations  that  are  then  evaluated  using  the 
objective  function  to  determine  the  impact  of  the  new  information  provided.  This 
process  is  repeated  several  times  for  each  supplemental  set  of  sample  sites 
considered.  Sets  of  sample  locations  that  demonstrate  only  small  modeling 
differences  from  one  realization  to  the  next,  or  those  indicating  significant 
differences  to  the  models,  but  where  the  differences  have  little  impact  on  the 
objective  function,  should  be  excluded  from  further  consideration. 


45 


This  resampling  procedure  can  be  very  expensive  in  terms  of  computer  time, 
especially  when  there  are  flow  and  transport  models  that  must  be  reevaluated  at 
each  step.  Consequently,  effective  procedures  are  required  for  selecting  sets  of 
sample  sites  for  consideration. 

What  is  Meant  by  an  Efficient  Optimization  Algorithm 

There  are  several  techniques  available  for  selecting  supplemental  sample  sites. 
Some  techniques  will  yield  an  exact  optimal  solution;  other,  less  computer 
intensive  techniques,  will  yield  a  "near  optimal"  solution.  The  distinction  between 
these  solutions,  and  the  considerations  that  might  provide  a  preferred  technique  for 
any  specific  application,  are  outlined  in  this  subsection. 

Spatial  sampling  problems  can  be  approached  as  a  combinatorial  optimization 
problem  where  a  large  but  finite  number  of  solutions  are  available.  Each  possible 
solution  (a  set  of  supplemental  sample  locations)  can  be  compared  through  the 
objective  function,  to  alternatives,  using  the  resampling  procedure. 

The  ideas  involved  in  the  sample  set  selection  algorithm  are  either  to  find  an 
exact  optimal  solution  dismissing  as  many  solutions  as  possible  without  a  formal 
evaluation  (recall  that  formal  evaluation  may  require  many  executions  of  a 
computer  intensive  simulation  involving  a  flow  and  transport  code),  or  to  find  a 
near  optimal  solution  using  the  fewest  comparisons  possible.  These  methods  can 
be,  and  probably  are,  most  effectively  used  in  combination. 

Simulated  annealing  and  genetic  algorithms  are  methods  that  compare 
solutions  iteratively,  providing  a  near  optimal  solution  where  "near"  depends  on 
the  efficiency  of  the  algorithm  and  the  number  of  computational  iterations 
involved.  Branch  and  bound  algorithms  can  be  used  to  find  an  exact  optimal 
solution.  These  algorithms  derive  their  benefit  (compared  to  an  exhaustive  search 
of  every  possible  solution  set)  by  dismissing  large  numbers  of  solutions  through  a 
single  objective  function  comparison.  Hybrid  methods  are  available  that  utilize 
features  of  all  of  these  random  search  methods. 

Details  concerning  the  concepts  and  methods  involved  in  these  algorithms  can 
be  found  in  references  cited  in  Rutherford  (1996).  Details  of  how  the  procedures 
might  best  be  implemented  for  spatial  sampling  or  monitoring  problems  can  also  be 
found  in  that  paper. 


46 


Chemical  Waste  Landfill  Example 

This  example  is  provided  to  illustrate  how  these  spatial  sampling  procedures 
might  work.  The  procedure  demonstrated  here  was  implemented  using  a  spatial 
sample  tool  (OPTMAS),  developed  at  Sandia  National  Laboratories,  for  the 
purpose  of  locating  soil  sample  locations  to  provide  the  maximum  relevant  site 
information.  The  site  analyzed  is  a  chemical  waste  landfill  site  at  Kirtland  Air 
Force  Base  in  Albuquerque,  New  Mexico.  The  objective  of  this  analysis  (for 
purposes  of  illustration)  is  to  determine  the  extent  of  a  fixed  contamination 
contour  with  a  90  percent  level  of  certainty. 

Figure  3  illustrates  the  present  sample  locations  by  the  black  dots.  Clearly, 
previous  sampling  was  performed  according  to  some  systematic  prescription.  The 
center,  shape,  and  extent  of  the  site  is  presumed  to  be  unknown,  but  it  is  assumed 
that  the  contamination  occurs  throughout  a  contiguous  region  or,  at  least,  that  the 
contour  associated  with  the  outer  boundary  is  the  one  of  interest.  This  is  a  realistic 
problem,  for  example,  in  situations  where  the  concern  is  to  achieve  a  certain  level 
of  assurance  that  a  contaminant  is  contained  within  site  boundaries.  It  was  further 
assumed  that  15  additional  sample  locations  were  to  be  specified  at  each  iteration 
throughout  the  analysis.  (That  is,  15  sample  locations  are  required  here.  These 
samples  will  be  taken  and  analyzed  and  the  process  will  be  repeated  if  necessary). 


Figure  3 

The  approach  taken  to  estimate  the  extent  of  these  contours  is  the  approach 
described  in  previous  sections  of  this  paper,  but  with  a  number  of  simplifications. 
First,  the  simulated  realizations  are  generated  using  sequential  Gaussian  simulation 
(because  it  was  readily  available,  but  not  necessarily  the  best  approach  to  modeling 
a  contaminated  region).  The  set  of  realizations  was  used  to  compute  the  pointwise 
variance  as  illustrated  in  Figure  3  (the  darker  areas  indicate  higher  variance).  The 


47 


reduction  in  variance  is  the  assumed  objective  function  criterion  for  this  example. 
Therefore,  after  the  pointwise  variance  of  the  set  of  realizations  has  been 
determined,  the  realizations  are  no  longer  needed  for  the  remainder  of  the  analysis 
(until  the  next  sampling  iteration).  This  further  simplication  is  possible  because  it 
is  assumed  for  simplicity  that  the  variance  is  reduced  at,  and  around,  sample 
locations  in  a  way  proportional  to  the  reduction  indicated  by  the  previously 
simulated  realizations.  The  details  are  provided  in  Rutherford  (1996).  This  is  a 
luxury  that  is  not  offered  in  the  generic  spatial  sampling  problem  where  new 
(hypothetical)  realizations  are  required  to  evaluate  each  set  of  potential  sample 
locations.  Finally,  it  was  assumed  for  reasons  outside  the  scope  of  this  paper  that 
the  sampling  had  to  occur  outside  the  10th  percentile  bound  for  this  contamination 
contour. 

Alternative  sample  locations  were  selected  using  the  hybrid  sample  selection 
algorithm  mentioned  above,  and  detailed  in  Rutherford  (1996).  For  each  sample 
set  selected  in  the  simulated  annealing  or  genetic  algorithm  portions  of  the 
procedure,  the  objective  function  (total  variance  throughout  the  region)  was 
computed.  The  sample  locations  were  dismissed  or  retained  as  a  candidate 
solution  for  the  problem  as  prescribed  by  the  hybrid  sample  site  selection 
algorithm.  The  circled  locations  in  Figure  3  present  a  near  optimal  solution  to  this 
problem.  It  is  difficult  to  determine  exactly  how  close  they  are  to  the  true  optimal 
solution  without  using  an  exact  approach.  For  the  present  illustration,  it  suffices  to 
say  that  the  solution  appears  to  satisfy  the  obvious  desirable  characteristics  of  a 
solution  to  the  problem  —  the  samples  are  in  areas  (constrained  as  described 
earlier)  where  the  variability  is  relatively  high  and  they  are  spread  out  throughout 
this  region  so  that  the  reduction  in  variance  achieved  by  sampling  at  one  location 
will  have  a  minimal  effect  on  the  variance  reduction  achieved  by  sampling  at 
another  selected  location. 

This  simplified  example  was  chosen  with  a  relatively  intuitive  solution  to 
illustrate  the  process.  The  samples  selected  in  a  more  realistic  example  might  be 
more  difficult  to  analyze  intuitively. 

References 

Mc  Kinney  DC,  and  D.P.  Loucks  (1992).  "Network  Design  for  Predicting 
Groundwater  Contamination",  Water  Resources  Research,  Vol.  28,  No.  1. 

Rutherford,  B.  (1996).  "A  General  Approach  to  the  Problem  of  Selecting  Optimal 
Spatial  Sampling  or  Monitoring  Locations",  white  paper  (in  progress),  Sandia 
National  Laboratories,  Albuquerque,  NM  87106-0829 


48 


Consideration  of  Spatial  Variability 
in  the  Management  of  Non-Point 
Source  Pollution  to  Groundwater 

W.  Woldt1,  F.  Goderya2,  M.  Dahab3,  and  L  BogardP 

Abstract.  —  Geostatistical  simulation  and  unsaturated  zone  modeling  are 
combined  to  evaluate  the  impact  of  spatial  variability  of  selected 
parameters  on  groundwater  nitrate  contamination  from  agricultural 
production.  Three  management  scenarios  involving  spatially  variable 
application  were  investigated  with  consideration  of  spatial  variability  in 
residual  soil  nitrate,  yield,  and  hydraulic  conductivity.  The  process  is 
applied  to  three  conceivable  situations,  differing  in  the  extent  of  spatial 
variability.  One  decision  input,  the  fertilizer  amount,  is  distinct  for 
different  scenarios  as  well  as  spatial  location  for  one  scenario.  Modeling 
results  indicate  that  variable  application  of  nitrogen,  based  on  spatially 
variable  parameters,  not  only  reduced  the  over-application  of  nitrogen, 
but  also  reduced  the  overall  groundwater  contamination  potential.  The 
modeling  results  indicate  that  spatial  management  techniques  hold 
promise  for  maintaining  production  while  simultaneously  protecting  the 
environment. 


INTRODUCTION 

The  natural  and  management  induced  variabilities  of  field  soils  and  hydrologic 
formations  are  extensively  acknowledged  as  dominant  factors  influencing  fluid 
and  mass  transport  through  the  subsurface  zone.  The  variations  of  soil  and  crop 
properties  have  led  to  attempts  to  understand  those  variations  and  to  manage 
production  accordingly.  The  management  of  production  according  to  localized 
conditions  is  variously  known  as  spatially  variable,  site  specific,  soil  specific, 
precision,  or  prescription  production.  A  consensus  does  not  appear  to  have  been 
reached  within  the  scientific  community  on  the  proper  term  or  definition. 
However,  the  methodology  presented  in  this  research  complements  production  and 
environmental  concerns  according  to  localized  conditions  and  will  be  referred  to 
as  spatial  management  technology. 


Assistant  Professor,  Department  of  Biological  Systems  Engineering,  University  of  Nebraska  -  Lincoln,  Lincoln  NE. 

2  Graduate  Research  Assistant,  Department  of  Gvil  Engineering,  University  of  Nebraska-Lincoln,  Lincoln,  NE. 

3  Professor,  Department  of  Civil  Engineering,  University  of  Nebraska-Lincoln,  NE. 


49 


The  key  factor  driving  spatially-variable  control  is  field  variability.  Spatial 
variability  may  be  random  or  auto-correlated,  may  be  long  distance  or  short 
distance,  and  may  be  small  or  large.  However,  all  affect  the  feasibility  of 
spatially-variable  control  and  the  design  of  a  particular  system  to  achieve  such 
control.  In  addition,  variability  may  result  from  both  natural  processes  and 
management  practices.  If  there  were  no  variability,  production  through  traditional 
practice  with  proper  adjustment  for  field  conditions  would  be  adequate.  For  all 
its  importance,  spatial  variability  and  it's  impact  on  crop  production  and  the 
environment  has  not  been  studied  to  a  large  extent.  Furthermore,  the  procedures 
for  developing  an  understanding  of  the  variability  in  a  particular  field  and 
effective  utilization  of  that  information  to  reduce  NPS  groundwater  pollution, 
while  at  the  same  time  maintaining  production  are  still  not  clearly  defined.  The 
purpose  of  this  research  is  to  develop  a  framework  for  exploring  the  effect  of 
spatial  variability  related  to  different  field  scale  parameters  in  managing  crop 
production  and  minimizing  groundwater  contamination. 

METHODOLOGY 

The  methodology  outlined  in  this  research  incorporates  spatial  variability  of 
various  parameters  on  a  field  scale  in  estimating  nitrate  contamination  to 
groundwater.  Unsaturated  zone  modeling  in  combination  with  conventional 
measurements  and  simulation  is  utilized  to  quantify  the  effect  of  field  spatial 
variability  on  contaminant  loading  for  quantifying  environmental  benefits  of 
spatial  management  technology.  The  approach  involves  geostatistical  simulation 
to  generate  a  number  of  realizations  reflecting  differing  degrees  of  spatial 
variation.  The  model  is  used  simulate  crop  growth  and  maintain  a  nitrogen  mass 
balance  in  the  system.  Primary  elements  of  the  methodology  are  discussed 
below. 

Management  Scenarios  and  Variability 

Three  application  scenarios  were  developed  in  the  methodology  leading  to  the 
spatial  management  technology.  The  spatial  management  scenario  is  based  on 
appropriate  management  of  nitrogen  to  meet,  but  not  to  exceed  production  needs. 
Phosphorus  and  other  nutrients  are  not  assumed  to  be  limiting  factors.  The 
scenarios  include  traditional  practices  and  are  modified  to  reflect  possible 
advancement  through  application  of  spatial  management  technology  (Figure  1). 

The  first  scenario  is  designed  to  consider  uniform  input  of  nitrogen  based  on 
traditional  practices  with  the  input  rate  set  at  a  "typical"  level  and  held  constant 
through  time.  The  second  scenario  assumes  the  same  crop  and  uniform 
application,  but  the  amount  of  application  is  modified  based  on  soil  and  yield 
information  from  a  single  location  in  the  field.  Thus,  the  application  is  uniform 
over  the  field,  but  is  variable  with  respect  to  time. 


50 


Scenario  1  Scenario  2  Scenario  3 


UA-TP  UA-l  CP  VA-16CP 


Legend: 

UAWT  Unifonn  Application  over  Time     UA-TP  Uniform  Application,  Traditional  Practice 
VAWT  Variable  Application  over  Time     UA-l  CP  Unifonn  Application,  1  Control  Point 

VA-16  CP  Variable  Application,  16  Control  Point 

Figure  1.  -  Graphical  illustration  of  three  application  scenarios. 

The  third  scenario  uses  variable  application  based  on  the  spatial  variation  of 
selected  model  and  decision  parameters.  The  selected  parameters  include  residual 
soil  nitrate,  crop  yield  and  soil  hydraulic  conductivity.  The  field  is  divided  into 
sixteen  sectors,  and  the  application  rate  is  based  on  a  single  hypothetical 
measurement  of  the  decision  variables  in  each  sector.  Hence,  the  field  will 
receive  an  application  with  respect  to  position  as  well  as  time  conforming  to 
decision  variables  within  the  sector. 

The  methodology  is  applied  to  three  different  cases  which  are  distinguished  by 
the  magnitude  of  field  variability.  They  are  referred  to  hereafter  as;  low, 
medium,  and  high  variability  case.  In  the  low  variability  case,  available  field 
information  for  residual  soil  nitrate  and  yield  was  used  to  develop  the  lateral  and 
vertical  parameter  distributions  and  geostatistical  relations  (Goderya  et  al.,  1996; 
Goderya,  1996).  Field  conditions  were  then  generated  for  medium  and  high 
variability  cases.  Mean  values  for  these  fields  were  comparable  to  the  observed 
field  data.  However,  each  case  is  distinguished  by  differing  coefficients  of 
variation.  For  this  study,  values  obtained  from  the  literature  were  used  to  define 
these  coefficients  of  variation  (Goderya,  1996).  Using  the  subjective  prior  values 
of  mean,  standard  deviation  and  correlation  length  values  for  selected  parameters, 
a  number  of  equally  likely  unconditional  realizations  were  generated.  Employing 
this  approach,  medium  and  high  variability  field  conditions  were  generated  for 
residual  soil  nitrate  and  yield  values.  The  general  statistical  properties  for  low, 
medium  and  high  variability  fields  are  given  in  Table  1. 

Geostatistical  Simulation  and  Mass  Balance 

A  total  of  100  field  realizations,  each  realization  having  120  locations,  were 
simulated  for  each  identified  spatially  variable  parameter  under  each  variability 
case.  Simulated  fields  were  then  used  as  an  input  to  an  unsaturated  zone 
transport  model  for  predicting  crop  production  and  nitrate  loading  to  groundwater. 


51 


Table  1. —Generated  Held  properties  for  low,  medium,  high  variability  cases. 


Residual  N03  (kg/ha) 

Yield  (kg/ha) 

Hyd.  Cond.  (mm/hr) 

Low 

Medium 

High 

Low 

Medium 

High 

Low 

Medium 

High 

Mean 

165.3 

146.9 

153.9 

8965 

8580 

8378 

2.37 

2.28 

1.29 

Std.  Dev. 

51.5 

88.5 

163.5 

815 

2350 

3460 

0.65 

2.08 

1.74 

Coeff.Var. 

0.31 

0.60 

1.06 

0.09 

0.27 

0.42 

0.27 

0.91 

1.35 

Minimum 

65.7 

29.4 

5.2 

6300 

4156 

2112 

0.32 

0.04 

0.01 

Maximum 

362.5 

518.8 

861.0 

10655 

12535 

14100 

6.69 

9.15 

9.99 

Hence,  the  complete  process  of  geostatistical  simulation  and  unsaturated  zone 
transport  modeling  was  executed  within  each  variability  case  for  each  spatial 
management  scenario.  The  spatial  input  data  changed  between  each  set  of  field 
realizations,  but  remained  constant  between  each  spatial  management  scenario. 
On  the  other  hand,  decision  input  changed  between  each  spatial  management 
scenario,  but  remained  identical  for  various  field  realizations.  For  example, 
spatial  inputs  of  residual  soil  nitrate,  yield,  and  hydraulic  conductivity  were 
different  between  realizations  1  and  2,  but  they  were  identical  between  realization 
1  of  scenario  1  and  realization  1  of  scenario  2  and  3. 

A  quasi-three-dimensional  approach  to  modeling  the  important  processes  and 
maintaining  a  mass  balance  in  the  system  is  used  in  this  research.  In  this  case, 
a  one-dimensional  model  is  combined  with  geostatistical  simulation  to  represent 
the  heterogeneity  (spatial  variability)  in  a  typical  field.  The  intention  of  this 
formulation  is  to  investigate  contaminant  loading  and  production  from  a 
heterogeneous  field  within  a  spatial  framework.  The  model,  TDNIT,  is  used  in 
this  research  for  simulating  nitrate  movement  through  root-  and  vadose-  zone 
(Goderya  et  al.,  1995).  However,  the  methodology  is  not  dependent  on  the  type 
of  model.  The  use  of  the  selected  transport  model  offers  the  advantage  of  short 
computation  time  and  reduced  input  data  demands.  The  results  from  this  model 
also  compared  well  with  the  program  Erosion  Productivity  Impact  Calculator 
(EPIC)  in  terms  of  its  predictive  ability  (Goderya  et  al.,  1995). 

A  total  of  48  simulation  runs,  each  encompassing  100  field  realizations,  were 
completed  for  this  effort.  The  resulting  decision  input  and  output  were  analyzed 
for  each  individual  model  node,  and  for  each  run  over  a  five  year  period. 

CASE  STUDY 

A  case  study  is  presented  to  demonstrate  the  methodology.  A  farm  field  in 
Central  Nebraska  represents  a  typical  crop  production  area  with  the  potential  for 
nitrate  contamination  that  may  be  found  in  the  Midwest.  Site  selection  was  based 
upon  availability  of  data,  shallow  water  table,  highly  permeable  vadose  zone, 
rather  uniform  soil  characteristics,  and  homogeneity  of  crops  produced  as  well  as 
the  general  agricultural  practices  employed.  Here,  most  of  the  land  adjacent  to 


52 


the  site  is  cropland  and  nitrogen  fertilizer  is  a  major  source  of  crop  nutrients  and 
nitrate  contamination  to  groundwater.  Continuous  corn  is  selected  as  the  crop  to 
be  simulated  since  it  is  the  major  crop  grown  in  the  case  study  area. 

Site  specific  information  for  the  model  input  variables  was  compiled  from  a 
variety  of  sources.  The  transport  program  required  meteorological  input  data  in 
the  form  of  daily  precipitation  (including  irrigation  amounts),  average  monthly 
temperature,  average  monthly  solar  radiation  and  albedo.  Additional  input  data 
included  depth  to  groundwater,  potential  mineralization  amounts,  the  number  of 
soil  layers,  soil  data  for  each  layer,  plant  data,  and  the  initial  moisture  conditions 
for  each  layer.  The  soil  input  data  (porosity,  field  capacity,  permanent  wilting 
point,  sand  content,  organic  matter  content,  saturated  hydraulic  conductivity,  and 
residual  water  content)  and  initial  conditions  (water  content  and  nitrate  content) 
for  the  field  were  based  on  the  vertical  profile  of  the  soil  horizon  and  the 
unsaturated  zone.  The  plant  data  include  crop  type,  four  different  nitrate  uptake 
coefficients,  leaf  area  index  values  and  root  depth  values  as  a  function  of  time, 
dry  matter/yield  ratio,  and  potential  maximum  yield.  The  nitrogen  fertilizer  data 
include  application  amount,  date  and  depth  of  application. 

RESULTS  AND  DISCUSSION 

The  model  data  were  evaluated  in  terms  of  area- wide  leaching  potential  and 
subsequently  for  leaching  as  a  function  of  spatial  management  practice.  For  all 
evaluations,  output  variables  were  used  for  relative  comparisons.  These  results 
represent  short-term  impacts  of  the  spatial  management  technology  and  were 
calculated  for  the  following  variables  of  interest:  annual  fertilizer  input,  annual 
crop  uptake  of  nitrogen,  and  nitrogen  losses  to  groundwater. 

Figure  2  summarizes  the  primary  nitrogen  input.  In  the  low  variability  case, 
nitrogen  fertilizer  input  was  set  at  200  kg/ha/yr  (178  bu/ac/yr)  for  the  traditional 
practice  scenario,  which  compares  favorably  with  the  average  corn  fertilization 
rate  in  Nebraska  (Follet  et  al.,  1991).  Fertilizer  nitrogen  inputs  were  lower  for 
second  scenario  because  of  the  information  contained  in  the  field  sample  which 
was  used  to  adjust  application  rate.  The  overall  amount  of  nitrogen  fertilizer 
applied  to  the  crop  in  the  third  scenario  was  reduced  further  due  to  the  added 
level  of  information  and  use  of  spatial  application  methods. 

The  simulated  crop  nitrogen  uptake  was  used  as  a  measure  of  yield  potential. 
The  annual  uptake  was  similar  under  management  scenarios  1  and  2  for  all  three 
cases  of  variability  (Figure  3).  However,  there  was  a  detectable  decrease  in 
uptake  for  scenario  3  (ie.,  spatial  management  technology).  Based  on  these 
results,  it  appears  as  though  production  based  on  spatial  management  technology 
compares  reasonably  well  with  the  traditional  baseline  output.  Hence,  while  the 
spatial  management  scenario  reduced  the  fertilizer  requirements,  it  also  resulted 
in  simulated  reduction  in  the  uptake. 

The  predicted  amount  of  nitrate-nitrogen  loading  to  the  groundwater  is 
presented  in  Figure  4.   Potential  nitrate  loading  to  groundwater  is  significantly 


53 


_250 

CD 

2l:2oo 

.3  150 


.§  100 
^  50 


0 


L1  L2  L3       M1  M2  M3       H1  H2  H3 
Scenarios 


Figure  2.  -  Annual  nitrogen  input  for  variability  cases  and  management  scenarios. 


L1  L2  L3      M1  M2  M3      H1  H2  H3 

Scenarios 


Figure  3.  -  Simulated  annual  crop  uptake  for  variability  cases  and  management  scenarios. 


500 


400 

cd 

""cD^ 

300 

cn 

"-£5 

CD 
CD 

200 

1 

CD 

100 

0 


L1  L2  L3      M1  M2  M3      H1  H2  H3 

Scenarios 


Figure  4.  -  Simulated  nitrate  leaching  for  variability  cases  and  management  scenarios. 


54 


reduced  for  all  three  scenarios.  However,  the  reduction  is  especially  evident  in 
the  third  scenario  with  high  levels  of  field  variability.  The  results  of  these 
simulations  are  consistent  with  the  magnitude  of  the  responses  reported  from 
experimental  studies  in  the  Midwest  (Ferguson  et  al.,  1994;  Phillips  et  al.,  1993). 

These  results  suggest  that  increased  information  for  different  parameters  and 
conditions  in  a  spatial  management  system  may  not  necessarily  result  in  a 
significant  decrease  in  nitrate  loading  to  groundwater.  In  fact,  it  appears  that  the 
success  of  spatial  management  technology  is  highly  dependent  on  the  degree  of 
spatial  variability  of  the  primary  field  parameters  that  influence  nitrate  fate  and 
transport.  Thus,  the  effort  associated  with  spatially  variable  application  in  a  field 
with  inherent  low  variability  of  key  parameters  would  most  likely  be  too  costly 
to  justify  the  benefits.  However,  as  the  field  variability  of  the  key  process 
dependent  parameters  increases,  the  apparent  environmental  benefits  of  spatial 
management  technology  become  clearer  as  depicted  in  Figure  4. 

SUMMARY  AND  CONCLUSIONS 

The  methodology  presented  in  this  research  incorporates  spatial  variability  of 
various  parameters  in  estimating  groundwater  contamination  potential.  It  employs 
a  combination  of  geostatistical  simulation  and  unsaturated  zone  transport 
modeling.  The  methodology  outlined  in  this  research  was  applied  to  three 
conceivable  variability  cases,  differing  in  extent  of  spatial  variability.  In  all  the 
cases,  management  practices  of  spatially  variable  application  were  evaluated  on 
the  basis  of  sustaining  agricultural  production  and  minimizing  environmental 
pollution.  Three  management  scenarios,  including  one  using  spatial  management 
technology  were  developed  and  evaluated  given  the  spatial  distribution  of  residual 
soil  nitrates,  yield,  and  hydraulic  conductivity  in  the  fields. 

The  first  case  was  defined  as  one  exhibiting  low  variability.  Results  of 
scenario  modeling  for  this  case  indicate  that  while  there  is  a  slight  reduction  in 
nitrate  leaching,  the  use  of  spatial  management  technology  may  not  necessarily 
result  in  substantial  benefits.  It  appears  that  uniform  application,  given  realistic 
yield  goals  and  composite  soil  samples,  may  achieve  results  comparable  to  spatial 
management  technology  in  fields  exhibiting  low  spatial  variability.  However,  use 
of  variable  rate  application  not  only  reduced  the  over-application  of  nitrogen,  but 
also  reduced  the  non-point  source  pollution  to  groundwater  due  to  agricultural 
practices  for  fields  exhibiting  both  medium  and  high  variability.  The 
methodology  described  in  this  study  provides  a  framework  to  ascertain  if  the  use 
of  spatially  variable  fertilizer  application,  or  spatial  management  technology,  is 
environmentally  friendly  and  also  provides  a  basis  to  evaluate  the  economics  of 
the  technology. 

ACKNOWLEDGEMENTS 

This  paper  is  supported,  in  part,  by  the  Center  for  Infrastructure  Research,  the 
Water  Center,  and  the  Agricultural  Research  Division  at  the  University  of 


55 


Nebraska-Lincoln  and,  in  part,  by  the  Cooperative  State  Research  Service  (CSRS) 
of  the  U.S.  Department  of  Agriculture. 

REFERENCES 

Ferguson,  R.B.,  G.W.  Hergert,  J.E.  Cahoon,  T.A.  Peterson,  C.A.  Gotway,  and 
A.H.  Hartford,  1994,  "Managing  spatial  variability  with  furrow  irrigation 
to  increase  nitrogen  use  efficiency,"  Second  International  Conference  on 
Site-Specific  Management  for  Agricultural  Systems,  Bloomington  / 
Minneapolis,  Minnesota,  March. 

Follet,  R.F.,  D.R.  Keeney,  and  R.M.  Cruse,  1991,  "Managing  nitrogen  for 
groundwater  quality  and  farm  profitability,"  Published  by  Soil  Science 
Society  of  America  Inc.,  Madison,  Washington. 

Goderya,  F.S.,  1996,  "Evaluation  and  estimation  of  groundwater  potential  using 
spatial  parameters,"  Ph.D.  Dissertation,  Department  of  Civil  Engineering, 
University  of  Nebraska,  Lincoln,  Nebraska. 

Goderya,  F.S.,  M.F.  Dahab,  W.E.  Woldt,  I.  Bogardi,  1995,  "Comparison  of 
Two  Transport  Models  for  Predicting  Nitrates  in  Percolating  Water", 
Submitted  for  publication  in  Transaction  of  the  ASAE. 

Goderya,  F.S.,  M.F.  Dahab,  W.E.  Woldt,  I.  Bogardi,  1996,  "Spatial  Patterns 
Analysis  of  Field  Measured  Residual  Soil  Nitrate",  In  Geostatistics  for 
Environmental  and  Geotechnical  Applications,  ASTM  STP  1283,  R. 
Mohan  Srivastava,  Shahrokh  Rouhani,  Marc  V.  Cromer,  A.  Ivan  Johnson, 
Eds. ,  American  Society  for  Testing  and  Materials,  Philadelphia  (In  press). 

Phillips,  D.L.,  P.D.  Hardin,  V.W.  Benson,  and  J.V.  Baglio,  1993,  "Nonpoint 
source  pollution  impacts  of  alternative  agricultural  management  practices 
in  Illinois:  A  simulation  study",  J.  Soil  and  Water  Conservation, 
48:5:449-457. 

BIOGRAPHICAL  SKETCH 

Wayne  E.  Woldt  is  an  assistant  professor  in  the  Department  of  Biological  Systems 
Engineering  at  the  University  of  Nebraska-Lincoln.  His  interests  include  risk  analysis, 
spatial  management  technology,  and  environmental  systems  analysis. 

Farida  Goderya  is  a  graduate  research  assistant  in  the  Department  of  Civil  Engineering 
at  the  University  of  Nebraska-Lincoln.  She  is  currently  completing  her  Ph.D. 
dissertation  on  spatial  management  technology. 

Mohamed  F.  Dahab  is  a  professor  in  the  Departments  of  Civil  Engineering  and 
Biological  Systems  Engineering  at  the  University  of  Nebraska-Lincoln.  His  interests 
include  pollution  prevention  and  nitrate  treatment  methods  for  municipal  water  supplies. 

Istvan  Bogardi  is  a  professor  in  the  Department  of  Civil  Engineering  at  the  University 
of  Nebraska-Lincoln.  His  interests  include  risk  assessment,  systems  analysis  for 
environmental  issues,  and  global  climate  change. 


56 


Incorporating  Soil  Variability  into  a 
Spatially  Distributed  Model  of 
Percolate  Accounting 

Andrew  S.  Rogowski1 


Abstract.-The  Soil-Plant-Atmosphere-Water  simulation  model  (SPAW)2 
was  used  to  compute  percolate  flux  below  the  root  zone  from  profile  water 
balance.  The  water  balance  was  based  on  the  prevailing  conditions  in  the 
atmosphere,  physical  variables  describing  water  status  of  soil,  and 
physiological  variables  related  to  a  stage  of  growth  and  water  use  by 
plants.  Although  measured,  or  estimated  at  a  point,  most  of  the  above 
variables  are  likely  to  vary  in  space  and  in  time.  The  study  addresses  this 
variability  and  describes  how  to  simulate  spatial  and  temporal 
distributions  of  hydrologic  variables  which  are  conditioned  by  a  set  of 
field  observations.  We  use  sequential  indicator  simulation  to  convert  water 
flow  evaluated  at  a  point  in  relatively  few  locations,  into  spatial  flow 
estimates.  The  study  describes  distribution  of  vertical  and  horizontal 
percolate  flux  in  time.  It  is  cast  in  a  GIS  format  at  a  catchment  scale. 


INTRODUCTION 

To  describe  behavior  of  natural  systems  we  often  rely  on  field  observations  and 
numerical  models.  Observations  are  generally  made  at  a  point,  while  models 
attempt  to  integrate  the  response  spatially  and  temporally.  The  field  observations 
considered  representative  of  an  area,  or  volume  of  soil,  serve  as  input  to  a  model. 
At  other  times  when  field  data  are  not  available,  modelers  use  published  soil 
surveys  to  extract  pertinent  input  variables  needed  to  run  a  model.  Such  variables 
are  assumed  to  apply  uniformly  to  an  entire  area  mapped  as  the  same  soil  polygon. 
In  truth  however,  the  variables  represent  observations  made  at  a  site,  which 
although  typical  for  a  given  soil,  may  be  far  removed  from  the  study 
area.  Moreover  soil  polygons  themselves  are  far  from  pure  and  contain  inclusions 
of  associated  materials  (Rogowski  and  Wolf,  1994).  The  objective  of  this  study 
was  to  incorporate  spatial  soil  variability  expressed  as  different  descriptions  of  the 
same  soil  into  a  simulation  model  of  percolate  accounting. 

1  Soil  Physicist,  USDA-ARS-PSWMRL,  University  Park,  PA. 

2  The  SPAW  model  was  developed  by  Dr.Keith  Saxton  USDA-ARS  Pullman,  WA. 


57 


METHODS  AND  MATERIALS 


Soil-Plant-Atmosphere-Water  (SPAW)  Model 

The  SPAW  simulation  model  (Saxton  et  al.,  1992)  provides  estimates  of 
evapotranspiration  (ET),  plant  water  stress  and  layer-by-layer  changes  in  soil  water 
content  and  percolate  flux  from  below  the  root  zone.  The  model  is  designed  to 
provide  a  daily,  one  dimensional  (vertical)  water  budget  at  a  point.  Actual 
measurements  of  ET,  soil  water  content  and  other  input  variables  may  be 
substituted  for  default  values,  or  periodically  used  to  validate,  or  recalibrate  the 
model.  The  model  requires  climatic  and  plant  data  and  a  detailed  in  situ  horizon  by 
horizon  description  of  the  soil  profile,  including  pertinent  soil  water  characteristics, 
or  textural  classification  information  for  each  soil  layer. 

Potential  evapotranspiration  (PET)  is  determined  from  measured  values  of  pan 
evaporation.  The  PET  values  are  used  to  compute  actual  ET  which  is  then 
subtracted  from  the  available  soil  water  in  storage.  Daily  precipitation,  corrected 
for  interception  and  runoff,  infiltrates  and  is  distributed  among  soil  layers  by 
cascading  down  from  the  surface.  The  amount  of  water  which  reaches  the  bottom 
boundary  layer  becomes  a  deep  vertical  percolation  flux  (qv).  Adjustments  among 
layers  due  to  unsaturated  water  flow  by  redistribution  are  made  using  a  Darcy 
subroutine. 

The  output  consists  of  daily  summaries  of  soil  water  for  each  layer,  in  addition 
to  a  digital  and  graphical  output  of  precipitation,  ET,  and  percolation  flux  qv.  The 
major  feature  of  the  model  is  that  it  addresses  the  soil-plant-air  continuum  at  a 
point,  and  estimates  are  made  conditional  to  the  profile  hydrologic  properties, 
subject  to  the  local  climatic  and  biotic  factors. 

Study  Area  and  Soils 

The  study  area  described  in  this  application  is  on  the  Mahantango  Creek  watershed 
in  Pennsylvania,  USA,  (Rogowski  and  Wolf,  1994).  The  watershed  is  situated  in 
the  Valley  and  Ridge  physiographic  province  of  the  Appalachian  Mountains, 
characterized  by  varying  relief:  upland  hills,  valleys,  and  forested  mountain  ridges 
dissected  by  streams.  The  climate  is  humid  with  well-distributed  rainfall  (about 
1000  mm/yr.).  Elevation  ranges  from  less  then  100  m  in  the  valleys  to  over  500  m 
on  the  ridge  tops.  The  ridges,  valleys  and  streams  are  oriented  northeast- 
southwest,  corresponding  to  the  regional  strike  of  major  rock  formations,  while 
beds  dip  to  the  northwest  and  southeast  along  the  centrally  located  anticline.  The 
land  use  is  predominately  cropland  in  a  rotation  management  system  of  corn,  small 
grains,  and  meadow,  intermingled  with  numerous  tracts  of  woodland,  areas  of 
permanent  pasture  and  orchards. 


58 


Elevation  (m) 


<232 
232  -240 
240  -247 
247  -255 
255  -262 
262  -270 
277  -285 

>285 


1.2  km 


Aspect 
N 

NE  ■» 
NW  ■■ 

SE  ■■ 

SW 
W  ™ 


Slope  (%) 


0 
3 
8 
15 
25 


3 
8 
15 
25 
60 


Soils 

Ab  (Albrights) 
Bk  (Berks) 
Ca  (Calvin) 
Ht  (Hartleton) 
Ln  (LeckKill) 
Mk  (Mecksville) 
We  (Weikert) 
Wt  (Watson) 


Figure  1.  Overlay  maps  of  elevations,  aspect,  slope,  and  soils  based  on  the  detailed  5m 
DEM  of  the  study  area;  locations  of  conditioning  pedons  are  shown  on  the  soils'  map. 


59 


The  experimental  area  is  1.2  km  on  the  side  with  the  individual  cell  size  of  5x5  m 
corresponding  to  a  detailed  DEM.  Figure  1  shows  distributions  of  georeferenced 
(GPS)  elevations,  slopes,  aspects,  and  soils  at  the  experimental  site.  Points  on  the 
soils  map  indicate  assigned  locations  of  the  soil  pedons  used  as  input  to  the  SPAW 
model.  The  displays  and  calculations  were  carried  out  within  the  framework  of 
GIS  IDRISI3  (Eastman,  1993). 

Soil  Profiles 

Soil  profile  locations  {pedons)  were  assigned  at  random,  using  as  input  the 
information  from  PSU  Soils  Database  (Ciolkosz  and  Thurman,  1992).  This 
database  is  a  compilation  of  800  pedons  representing  170  different  soil  series 
which  have  been  collected  and  analyzed  by  the  soil  characterization  laboratory.  It 
contains  detailed  profile  and  site  descriptions,  as  well  as  layer-by-layer  physical, 
chemical  and  mineralogical  data.  There  are  generally  several  descriptions  of  the 
same  series  sampled  at  different  locations.  For  the  eight  soils  found  in  the 
experimental  area  60  pedons  were  described  in  the  database,  these  were 
augmented  with  pedons  of  a  different  series  considered  as  inclusions,  for  a  total  of 
100  pedons.  The  pedons,  assigned  at  random  to  soil  polygons  of  the  same  series 
(figure  1),  constituted  the  conditioning  data  sites. 

Conditioning  Data 

At  each  of  the  conditioning  sites  we  extracted  from  the  soils  database  the 
textural  information,  as  well  as  percent  sand,  silt,  and  clay,  and  depth  to  each  layer. 
We  used  daily  precipitation  and  PET  data  from  the  1990  season,  and  assumed  a 
uniform  crop  cover  (corn)  over  the  entire  area.  The  SPAW  model  was  run  on  each 
profile  from  Spring  to  Fall.  The  model  was  started  with  a  uniform  water  content 
(0V)  of  0.20  cm  /cm  and  recalibrated  to  volumetric  field  capacity  for  each  layer. 

Soil  Anisotropy 

Field  saturated  hydraulic  conductivity  Kfs  of  each  layer  was  estimated  from  the 
ten  textural  classes  in  the  SPAW  supplied  tables.  The  Kv  (vertical)  for  the  soil  was 
computed  as  a  harmonic  mean  of  component  layers  weighted  by  a  layer  thickness. 
The  Kh  (horizontal)  for  the  soil  was  taken  as  the  conductivity  of  top  layer,  if 
followed  by  a  layer  with  lower  K.  If  K  in  the  following  layer(s)  was  higher,  Kh  was 
taken  as  an  arithmetic  average  of  component  layers  weighted  by  the  layer 
thickness.  The  ratio,  n  =  Kh  /  Kv  was  assumed  to  represent  profile  anisotropy 
(Zaslawsky  and  Rogowski,  1969).  Given  percolation  flux  qv  computed  from  the 
SPAW,  horizontal  flow  component  qh  was  approximated  as, 

qh  =  qvx  n  x  tan  a  [1] 
where  a  represents  land  slope  in  %  (figure  1)  derived  from  a  detailed  5  m  DEM, 
and  tan  a  is  equivalent  to  (slope  gradient)"1  definition  in  IDRISI3 

3The  mention  of  tradenames  does  not  constitute  an  endorsement  of  the  product  by  the  US  Department  of  Agriculture. 


60 


Sequential  Indicator  Simulation 


Results  in  literature  (Journel  and  Alabert,  1989),  suggest  that  a  geostatistical 
conditional  simulation  approach  known  as  sequential  indicator  simulation  (SISIM) 
can  be  used  to  reproduce  spatial  patterns  of  continuity  and  to  introduce  a  measure 
of  uncertainty  into  soil  survey  data.  The  purpose  of  the  simulation  is  to  construct 
from  the  conditioning  data  (100  profiles),  numerous,  equiprobable  spatial 
"realizations",  or  outcomes,  of  an  attribute  qv,  that  would  reproduce  spatial 
patterns  associated  with  the  occurrence  of  extreme  values. 

To  model  different  portions  of  an  irregular  distribution,  the  sequential  indicator 
simulation  partitions  observed  population  into  several  classes  each  separated  by  a 
threshold.  For  each  observed  value  it  then  defines  the  indicator  transformation 
with  respect  to  that  threshold.  Thus  for  a  random  variable  Z(u)  at  a  location  u  and 
a  threshold  value  z,  the  corresponding  indicator  transformation  I(u;z)  would  be, 

I(u;z)=l,  ifZ(u)<z  [2] 
=  0,  otherwise 

Given  a  set  of  qv  values,  the  SISIM  procedure  converts  them  into  indicator 
distributions  for  each  threshold.  It  then  attempts  to  define  an  algorithm  for  adding 
an  indicator  value  at  an  unsampled  location  which  is  consistent  with  a  spatial 
structure  (variogram  or  covariance)  of  observed  population  for  a  particular 
threshold  (Deutsch  and  Journel,  1992). The  first  step  involves  estimation  of  the 
conditional  probability  that  a  new  value  is  less  then  a  given  threshold.  This  is  done 
by  kriging  each  unsampled  location  using  the  surrounding  indicator  values.  The 
resulting  estimate  of  the  conditional  cumulative  distribution  function  (ccdf)  is 
between  0  and  1.  The  actual  simulation  of  a  corresponding  indicator  value  at  an 
unsampled  grid  location  is  accomplished  by  drawing  a  random  number  between  0 
and  1  from  a  uniform  distribution  (Monte  Carlo).  If  this  number  is  less  than,  or 
equal  to  a  kriged  value,  a  simulated  indicator  value  of  1  is  assigned  to  that 
location.  If  it  is  more,  a  value  of  0  is  assigned.  The  new  value  becomes  a  part  of 
the  conditioning  data  set  and  the  procedure  is  repeated  at  another  location  until  all 
unsampled  locations  in  a  catchment  grid  are  filled. 

RESULTS 

Distributed  Flow 

After  generating  25  equiprobable  realizations  of  cumulative  percolate  flux  qv, 
expected  values  and  quantiles  of  the  average  of  25  distributions  were  analyzed  for 
three  dates:  March  30th,  August  15th,  and  November  15th,  1990.  We  show  these  in 
figure  2.  We  also  generated  25  realizations  of  n  =  Kh/Kv.  Subsequently,  expected 
values  of  //  and  qv  at  each  pixel,  and  slope  from  the  overlay  of  slope  in  figure  1 
were  used  in  Eq.  3  to  compute  distributions  of  the  horizontal  flux  component  qh 


61 


q»  March  30  q»  August  15 


1        2  3  4  1        2  3  4 

<1  IB  1.18-1.44  1  44-1.80  >1  80  <2.02  2.02-2.44  2.44-2.93  >2.93 


q,  November  1 5  qh  August  15 


12  3  4 

<3.83  3.83-».24  4.24-4.75  >4.75 


12  3  4 

<0 28  028-047  0.47-0.78  >0.78 


12      3  4 
III 


Figure  2.  Simulated  (SISIM)  quantiles  of  expected  values  of  qv  (cm)  on  Mar  30th,  Aug  15th, 
and  Nov  15th  and  of  qh  (cm)  on  Aug  15th  based  on  100  conditioning  data. 


On 


Probes  SI. 52  Prob.s23.72 


Figure  3  Distribution  of  qv  (cm)  for  three  quantiles  (Q)  of  25  SISIM  simulations  on  Augl5  ; 
and  probability  (Prob)  of  values  lower  than  Q5  (1.52  cm),  or  higher  than  Q95  (3.72  cm). 


62 


for  each  location  and  date.  Since  there  was  not  much  difference  among  the  three 
dates,  figure  2  shows  qh  for  August  15th  only.  We  represent  the  qv  and  qh  as 
quartiles  (Q)  of  the  average  (25  SIS)  distribution:  (1)  yellow  <  Qu  (2)  green  Qi  to 
Q2,  (3)  blue  Q2  to  Q3,  and  (4)  red  >Q3.  Results  suggest  that  areas  with  high  qv 
(red)  may  be  the  sites  of  preferential  recharge.  Similarly  locations  with  above 
average  qh  may  indicate  preferential  interflow,  or  runoff  producing  zones.  The 
values  of  qh  are  very  dependent  on  interaction  among  the  distributions  of  slope 
(from  5  m  DEM),  and  distributions  of  n  and  qv(from  two  separate  25  SISIM  runs). 

The  overall  distribution  of  qv  does  not  differ  much  from  season  to  season.  In  the 
Spring  a  very  prominent  red  zone  occurs  the  middle  of  the  catchment.  As  time 
goes  on,  this  zone  becomes  more  diffuse  in  the  Summer,  and  brakes  into  several 
patches  in  the  Fall.  When  results  for  qv  in  figure  2  are  compared  with 
corresponding  terrain  properties  in  figure  1,  the  primary  recharge  area  seems  to  be 
in  the  center  of  the  catchment.  This  is  also  zone  of  terrain  depression  and  a  natural 
flow-pathway  which  connects  part  of  the  catchment  to  the  main  drainage. 
Ironically,  the  farm  buildings  are  located  in  this  same  general  area. 

Quantile  Maps 

To  derive  a  quantile  map,  25  simulated  values  for  each  pixel  were  ranked  and  the 
quantiles  (Q),  were  selected  as  simulated  pixel  values  corresponding  to  the  25th, 
50th,  and  75th  percentile  in  the  ranking.  These  pixels  were  subsequently  combined 
into  a  single  quantile  overlay.  The  three  quantiles  (figure  3)  were  Q25  ,  median 
(Q50),  and  Q75.  The  Q25  realization  focuses  on  highlighting  the  potential  outliers, 
since  the  probability  of  a  lower  value  than  shown  is  only  25%.  The  median  (Q50) 
quantile  map  tends  to  minimize  the  mean  of  absolute  deviations  for  each  pixel 
location,  that  is,  individual  pixel  values  are  close  to  what  they  appear  to  be.  In  the 
Q75  map  the  focus  shifts  to  the  reliable  low  zones,  because  in  this  case  the 
probability  of  exceedance  would  be  only  25%.  This  map  and  the  Q50  map  are  the 
closest  in  appearance  to  that  in  figure  2.  The  results  suggest  that  the  red  zone  is 
real  and  its  occurrence  persists  in  the  quantile  maps.  Equally  persistent  are  the  low 
flow  (yellow)  areas  in  the  east  and  south  east  corner  and  along  the  lower  portion 
of  western  boundary. 

Probability  Maps 

Probability  maps  (figure  3)  provide  additional  detail  about  the  existence  and 
continuity  patterns  associated  with  either  the  low,  or  high  zones  of  qv.  The  maps 
are  expressed  as  percent  probability  that  an  area  is  either  higher,  or  lower  than  a 
given  cutoff.  The  choice  of  a  cutoff  is  arbitrary.  Our  maps  were  constructed  for 
two  cutoffs:  Prob  <  Q5  and  Prob  >  Q95.  The  P<Q5  map  delineates  potential  low 
flow  zones,  while  the  P>Q95  map  shows  high  flow  areas  in  terms  of  the  probability 
of  occurrence.  Both  maps  confirm  the  existence  of  centrally  located,  connected, 
high  qv  area  ,  and  corresponding  low  flow  zones  in  the  eastern  and  western 
portions  of  the  catchment. 


63 


CONCLUSIONS 


To  introduce  a  measure  of  spatial  variability  into  soil  survey  data,  we  extracted 
detailed  pedon  information  from  a  soils'  data  base  and  assigned  it  at  random  to 
locations  within  a  mapped  polygon  representing  that  soil.  The  assigned  pedon 
locations  became  the  sites  of  conditioning  data.  We  run  a  model  (SPAW)  to 
compute  daily  water  balance  for  each  pedon,  extracting  the  percolate  flux 
component  qv.  The  values  of  qv  were  then  used  as  conditioning  data  in  the 
sequential  indicator  simulation  model  to  distribute  qv  over  the  catchment.  Results 
suggested  the  existence  of  high  an  low  flow  zones  reflecting  terrain  properties. 
Their  continuity  was  confirmed  by  the  quantile  and  probability  of  exceedance 
maps.  An  added  benefit  was  the  identification  of  horizontal  flux  component  qh, 
related  to  soil  anisotropy.  Distributions  of  qh  were  strongly  slope  dependent  and 
reflected  contributions  by  the  percolate  flux  qv  and  local  anisotropy  n. 

ACKNOWLEDGMENTS 

Contributions  by  D.  Simmons  to  data  visualization  and  processing  are  appreciated. 

REFERENCES 

Ciolkosz,  E.J.  and  Thurman,  N.C.1992.  Soil  Char.  Lab.  Database  System.  Agron. 

Ser.  No.  124.  Agron.  Dept.,  Pen  State  Univ.,  University  Park,  PA  16802.  59p. 
Deutsch,  C.V.  and  Journel,  A.G.  1992.  GSLIB,  Geostatistical  software  library  and 

users  guide.  Oxford  Univ. Press,  New  York,  NY. 
Eastman,  J.R.  1992.  EDRISI  a  grid  based  geographic  analysis  system.  Version  4.1. 

Clark  University  Graduate  School  of  Geography,  Worcester,  MA  01610.  178p. 
Journel,  A.G  and  Alabert,  F.  1989.  Non-Gaussian  data  expansion  in  earth  sciences. 

Terra  Nova  1:  123-134. 
Rogowski,  A.S.  and  Wolf,  J.K.  1994.  Incorporating  variability  into  soil  map  unit 

delineations.  Soil  Sci.Soc.Am.J.,  58: 163-174. 
Saxton,  K.E.,  Porter,  M.A.,  and  McMahon    1992. Climatic  impacts  on  dryland 

winter  wheat  by  daily  soil  water  and  crop  stress  simulations.  Agric.  and  Forest 

Meteorology,  58:177-192. 
Zaslavsky,  D.  and  Rogowski,  A.S.  1969.  Hydrologic  and  morphologic  implications 

of  anisotropy  and  infiltration  in  soil  profile  development.  Soil  Sci.Soc.Am.J., 

33(4):  594-599. 

BIOGRAPHICAL  SKETCH 

Andy  Rogowski  is  a  Soil  Scientist  at  the  USDA-ARS  Pasture  Systems  and 
Watershed  Management  Research  Laboratory,  he  graduated  from  the  Iowa  State 
University  at  Ames  with  a  Ph.D.  degree  in  Soil  Physics. 


64 


Response  of  a  Watershed  Model  to 
Varying  Spatial  Landscape 
Characteristics 


Dennis  P.  Swaney1,  Wen  Ling  Kuo2,  David  A.  Weinstein3, 
Charles  Mohler4,  Stephen  DeGloria5,  Chris  Pelkie6, 
Fuan  Tsai7,  Tammo  S.  Steenhuis8,  and  Charles  E.  McCulloch9 


Abstract.  Using  a  spatially-explicit  deterministic  model  of  hydrology 
and  soil  nitrogen  dynamics  in  a  small  watershed  in  upstate  New  York, 
we  investigate  the  effect  of  varying  the  spatial  distribution  of  land  use 
on  processes  within  the  watershed.  Multiple  landscapes  having 
specific  properties  are  generated,  and  the  effects  on  runoff  of  water, 
runoff  loss  of  nitrogen,  and  other  soil  nitrogen  processes  are  observed 
over  a  period  of  two  years.  We  illustrate  techniques  of  visualization 
for  comparing  total  magnitude  and  spatial  and  temporal  distributions  of 
these  variables.  The  ultimate  goal  of  our  approach  is  to  develop  a 
methodology  for  use  in  comparative  studies  of  land  use  modification  at 
the  landscape  scale. 

INTRODUCTION 

Because  nitrogen  can  be  seen  both  as  a  pollutant  and  a  nutrient, 
understanding  the  nitrogen  dynamics  of  watersheds  is  crucial  to  both  surface 
and  groundwater  quality  management,  to  effective  agricultural  management, 
and  to  the  advancement  of  our  understanding  of  other  terrestrial  and  aquatic 
ecological  processes.  With  the  explosive  growth  of  computer  resources  over 
the  last  few  decades,  mathematical  modeling  of  nitrogen  at  the  watershed  scale 
has  advanced  from  empirical  loading  models,  to  process-based  lumped- 
parameter  models,  to  spatially  explicit  models.  While  the  first  two  categories 
of  model  permit  the  estimation  of  annual  or  seasonal  loading  rates,  and  the 
construction  of  average  mass  balances  of  some  watersheds,  they  fail  for 
complex  watersheds  in  which  flows  are  generated  primarily  in  small  "source 
areas".  They  also  do  not  allow  examination  of  the  impacts  of  the  position  of 

'  Extension  Associate.  Center  for  the  Environment,  Cornell  University,  Ithaca,  NY 
2  Graduate  Student,  Agricultural  and  Biological  Engineering,  Cornell  University 
'  Forest  Ecologist.  Boyce  Thompson  Institute  for  Plant  Research,  Ithaca,  NY 
4  Senior  Research  Associate,  Ecology  and  Systematics,  Cornell  University 
s  Associate  Professor,  Soil,  Crop  &  Atmospheric  Sciences,  Cornell  University 

''  Scientific  Visualization  Producer  .Center  for  Theory  and  Simulation  in  Science  and  Engineering,  Cornell  University 

Graduate  Student.  Civil  and  Environmental  Engineering,  Cornell  University 
K  Associate  Professor,Agricultural  and  Biological  Engineering,  Cornell  University 
'  Professor.  Biometrics  Unit  and  Statistics  Center.  Cornell  University 


65 


nitrogen  sources  in  the  watershed,  and  the  effects  of  correlations  between 
sources  and  topography  and  soil  properties  which  affect  nitrogen  transport  and 
transformations.  Even  some  of  the  most  successful  watershed  models  which 
incorporate  spatial  information  do  so  by  aggregating  topographic  or  soil  data 
(e.g.,  Beven  and  Kirkby,  1979)  or  by  using  monthly  or  greater  time  steps  (e.g., 
Bartell  and  Brenkert,  1990),  thereby  losing  the  detail  of  the  spatial  interactions 
and  the  response  to  individual  rainfall  events. 

Under  what  circumstances  should  these  details  matter?  If  we  are  considering 
alternative  strategies  for  managing  a  watershed  in  which  we  must  decide  the 
location  of  nitrogen  sources  (e.g.  cornfields)  it  may  be  of  interest  to  know 
whether  location  strongly  affects  nitrogen  load  from  the  watershed.  A  related 
question  is  whether  an  existing  mosaic  of  land  use  in  a  watershed  represents  the 
smallest  environmental  impact  in  terms  of  surface  or  groundwater  loading,  or  if 
alternative  mosaics  could  be  more  desirable. 

Below,  we  report  some  preliminary  results  of  our  analysis  of  a  small  upstate 
New  York  watershed  which  incorporates  data  describing  the  spatial  distribution 
of  topography,  land  use,  and  soils.  We  discuss  the  effects  of  changing  land  use 
characteristics  on  soil  water  and  nitrogen  spatial  distributions  and  losses  from 
the  watershed  in  runoff. 


METHODS 


Model  Descriptions 

Hydrology 

For  this  study,  we  simulate  the  hydrology  and  nitrogen  dynamics  of  a  15.3 
ha  watershed  in  upstate  New  York  on  a  daily  time  step  for  two  years, 
beginning  in  April,  1991.  We  use  a  GIS-based  hydrology  model  originally 
developed  by  Zollweg  and  Steenhuis  (Zollweg,  1994)  and  adapted  for  our 
application  (Kuo  et  al.,  1996)  to  drive  a  spatially  explicit  soil  nitrogen  model. 
Hydrology  is  driven  by  local  daily  precipitation  and  pan  evaporation  data 


precipitation  evapotranspiration,runoff 

x~^r — ; 


layer  1 


lateral 
interflows 


layer  2 


layer  3 


21 


layer  4 


Active 
root  zone 

Potential 
root  zone 


Sub-soil 


Saturated 
zone 


Figure  1.  Schematic  of  hydrology  model 


66 


obtained  from  the  Northeast  Regional  Climate  Center  at  Cornell.  The 
hydrology  model  describes  the  watershed  as  a  collection  of  1531  100  m2  cells 
which  are  divided  into  4  layers  (figure  1). 

The  top  layer  represents  the  active  root  zone  the  bottom  boundary  of  which 
varies  in  depth  through  the  growing  season.  Layer  2  is  the  zone  between  this 
depth  and  the  maximum  depth  of  the  root  zone.  Layer  3  represents  the 
unsaturated  zone  below  the  potential  root  zone,  and  is  bounded  below  by  the 
water  table.  Layer  4,  the  saturated  zone,  is  bounded  below  by  impervious 
bedrock,  the  depth  to  which  depends  on  local  soil  type. 

Water  enters  the  top  layer  of  each  cell  through  rainfall  and  snowmelt. 
Lateral  exchange  of  water  between  a  cell  and  its  eight  neighbors  is  determined 
by  the  product  of  local  hydraulic  conductivity  and  elevational  gradient. 
Vertical  exchanges  are  governed  by  mass  balance  between  layers,  given  the 
impervious  bottom  boundary  condition  of  the  bottom  layer.  Water  in  excess  of 
saturation  is  lost  from  the  top  layer  as  surface  runoff.  Water  is  also  lost  from 
the  active  root  zone  by  evapotranspiration. 

Nitrogen 

The  nitrogen  model  (figure  2)  includes  exchanges  between  three  components 
in  each  cell:  organic  N,  ammonia,  and  nitrate.  Nitrogen  is  added  to  the 
watershed  in  the  form  of  dry  deposition  (daily)  and  wet  deposition  (episodically 
with  rain)  of  nitrate  and  ammonium,  and  through  fertilizer  applications  of 
nitrate  on  agricultural  lands.  In  the  simulations  discussed  here,  a  100  kg/ha 
application  of  nitrate  to  crop  land  occurs  on  May  31  of  each  year.  Exchanges 
between  the  N  pools  are  mediated  by  microbial  activity  which  depends  on 
temperature  and  soil  moisture  content. 


Plant  deposition, 
deposition  uptake  Runoff 


lateral 

hydrologic 
transport 


Figure  2  .  Schematic  of  nitrogen  model 


67 


Organic  nitrogen  may  be  mineralized  to  ammonia,  which  in  turn  may  be 
nitrified  or  lost  to  the  atmosphere  through  volatilization.  Nitrate  is  lost  to  the 
atmosphere  by  denitrification  under  wet  soil  conditions.  Nitrate  is  assumed  to 
be  the  only  mobile  component  of  nitrogen.  It  moves  in  dissolved  form  with 
water  within  the  watershed,  and  is  lost  from  the  watershed  in  runoff  water. 
Both  ammonium  and  nitrate  are  taken  up  by  plants  following  a  demand 
schedule  determined  by  the  land  use  category  of  the  particular  cell. 

Landscape 

Land  use  data  characterizing  the  watershed  were  determined  from  1991  aerial 
photographs  (1:24,000;  figure  3a);  digital  elevation  data  were  digitized  from 
contours  of  USGS  1:24,000  topographic  maps  (figure  3c);  soil  maps  (1:20,000) 
from  the  Tompkins  County  soil  survey  (figure  3d)  (Poiani  et  al.  ,  submitted). 

We  consider  three  alternative  land  use  mosaics  for  the  watershed:  existing 
land  use  ("original"),  100%  active  cropland  ("agricultural")  and  a  "synthetic 
landscape"  generated  using  a  conditionally  independent  hierarchical  model 
(Searle,  et  al.,  1992)  ("random").  The  existing  land  use  mosaic  includes  five 
categories:  active  agriculture  (43%),  forest  (41%),  wetland  (9%),  inactive 
agriculture  (4%),  and  low  density  residential  (3%)  (figure  3a).  The  agricultural 
mosaic  was  included  to  examine  the  limiting  case  in  which  a  nitrogen  fertilizer 
is  applied  to  the  entire  landscape.  Detailed  description  of  the  synthetic 
landscape  generator  is  beyond  the  scope  of  this  paper,  but  is  provided  in 
McCulloch  et  al.  (1995).  Briefly,  the  land  use  categories  are  modelled  as  a 
function  of  slope,  and  are  generated  using  a  spatially  autoregressive  correlation 
structure  to  induce  a  high  correlation  between  land  uses  in  adjacent  cells. 
Synthetic  landscapes  are  useful  for  assessing  variation  and  for  "filling  in" 
landscapes  which  may  undergo  change.  The  example  realization  of  the 
synthetic  landscape  model  considered  here  resulted  in  the  following  distribution 
of  land  uses:  active  agriculture  (54%),  forest  (22%),  wetland  (6%),  inactive 
agriculture  (1%),  and  low  density  residential  (18%)  (figure  3b). 

RESULTS  AND  DISCUSSION 

Hydrology  and  nitrogen  were  simulated  for  two  years  beginning  April  1 . 
Here,  we  compare  the  water  and  nitrate  conditions  in  the  landscape  at  the  time 
of  the  fertilizer  application  to  cropland  in  the  second  year,  after  any  effects 
associated  with  the  initial  condition  (saturated  soil  water  content;  constant  N 
concentrations)  have  dissipated,  and  again  just  prior  to  the  start  of  the  next 
growing  season. 

Maps  of  the  total  soil  moisture  content  for  the  "original"  mosaic  are  shown 
in  figure  3e  and  3f.  We  show  the  results  from  this  case  only  because  those 
corresponding  to  the  other  mosaics  are  nearly  indistinguishable  from  this  one. 
The  figures  indicate  that  by  the  end  of  May  in  the  second  year  of  the 
simulation,  and  persisting  through  and  after  the  growing  season,  the  watershed 


68 


has  established  a  distinctive  pattern  of  high  moisture  content  along  a  central 
drainage  which  curves  to  the  lower  left  where  the  outlet  is  located.  Even 
during  the  driest  periods  of  the  simulation,  this  central  feature  exists,  expanding 
and  contracting  following  rainfall  and  snowmelt  events. 

The  average  nitrate  concentration  in  the  root  zone  corresponding  to  the  same 
dates  for  each  of  the  land  use  mosaics  are  shown  in  figures  4a  and  4b,  along 
with  cumulative  runoff  losses  of  nitrate.  For  all  three  landscape  mosaics,  the 
pattern  of  nitrate  concentration  follows  the  distribution  of  crop  land 
immediately  following  fertilizer  application,  as  expected,  because  the  rate  of 
nitrate  applied  is  at  least  an  order  of  magnitude  greater  than  that  from  typical 
atmospheric  deposition.  However,  about  nine  months  after  application,  the 
pattern  shows  the  effect  of  water  content  in  the  root  zone.  In  particular,  the 
wet  areas  that  occupy  the  center  of  the  map  and  curve  to  the  lower  left  appear 
to  be  largely  depleted  of  nitrate.  This  may  be  due  to  a  combination  of 
denitrification  and  runoff  losses,  both  of  which  occur  at  their  highest  rates 
under  saturated  conditions.  We  believe  the  transitional  pattern  from  a  land  use 
dominated  mode  immediately  following  a  large  application  of  nitrogen  to  a 
soil/topography  based  pattern  as  hydrological  processes  proceed  is  a  general 
feature  of  mixed  use  watersheds. 

Not  surprisingly,  cumulative  runoff  losses  exhibit  a  sharp  increase 
immediately  following  fertilizer  application  on  all  landscapes,  with  a  gradually 
increasing  trend  thereafter.  The  highest  losses  occur  from  the  "agricultural" 
mosaic.  The  "random"  mosaic  generates  nitrate  losses  intermediate  between  the 
"agricultural"  and  the  "original"  mosaic. 

Of  some  interest  is  the  relatively  high  loss  of  nitrogen  in  runoff  from  the 
synthetic  landscape.  High  losses  are  also  observed  in  two  alternative 
realizations  of  the  synthetic  landscape  not  shown  here.  Expressed  per  area  of 
crop  land,  the  losses  from  the  synthetic  landscapes  are  near  or  above  those 
generated  from  the  "agricultural"  mosaic.  One  interpretation  of  this  result  is 
that  a  more  sophisticated  landscape  generation  algorithm  should  take  persistent 
features  of  soil  moisture  (or  its  governing  variables:  topography,  porosity  and 
hydraulic  conductivity)  into  account.  It  seems  unlikely  that  the  factors  which 
have  selected  the  location  of  cropland  in  a  landscape  reduce  the  runoff  losses  of 
nitrogen,  but  further  work  with  synthetic  landscapes  may  shed  light  on  this 
issue. 

ACKNOWLEDGEMENTS 

Visualizations  of  the  simulation  results  reported  here  were  obtained  using  the 
resources  of  the  Cornell  Center  for  Theory  and  Simulation  in  Science  and 
Engineering.  The  Cornell  Theory  Center  receives  major  funding  from  the 
National  Science  Foundation  and  New  York  State,  with  additional  support  from 
the  Advanced  Research  Projects  Agency,  the  National  Center  for  Research 
Resources  at  the  National  Institutes  of  Health,  IBM  Corporation  and  members 
of  the  Corporate  Partnership  Program. 


69 


(a 


12.3 

0*2  i 


0.0™ 

CONCENTRATION 

N03  (mg/l) 


250m 


NO 3  in  roof  zone 
Ddy:  pH^T 


0.0 

CONCENTRATION 

N03  (mg/l) 


NO3  in  root  zone 


r»  inn; 

Day:      f  U  C 


Apr    Yr  1    Apr    Yr  2  Apr 


*• 


250m 


50 
25 


CUMULATIVE  NO3  RUNOFF  (kg) 

Apr    Yr  1    Apr    Yr  2  Apr 


Figure  4  (a)  Distribution  of  nitrate  nitrogen  in  the  root  zone  on  the 
date  of  the  second  fertilizer  application  (day  426)  and  corresponding 
cumulative  nitrate  losses  in  runoff  for  the  three  land  use  mosaics; 
(b)  same  variables  276  days  later. 


70 


Figure  3  (a)  Original  land  use  map.  (b)  "random"  land  use  map. 
(c)  elevation  map  with  original  land  use  map  superimposed  and  5' 
elevation  contour  lines,  (d)  soil  map.  (e)  moisture  saturation  index  at 
time  of  second  fertilizer  application,  (f)  saturation  index  276  da\s  later 


71 


BIOGRAPHICAL  SKETCHES 


D.  Swaney  is  a  mathematical  modeler  in  the  Biogeochemistry  Program  of  the 
Center  for  the  Environment  at  Cornell.  W.  Kuo  is  a  Ph.d.  candidate  in  the 
Dept  of  Agricultural  and  Biological  Engineering.  D.  Weinstein  is  a  forest 
ecologist  at  the  Boyce  Thompson  Institute  for  Plant  Research.  C.  Mohler  is  a 
senior  research  associate  in  the  Section  of  Ecology  and  Systematics  at  Cornell. 
S.  DeGloria  is  an  associate  professor  in  the  Dept  of  Soil,  Crop  and 
Atmospheric  Sciences.  C.  Pelkie  is  scientific  visualization  producer  at  the 
Cornell  Theory  Center.  F.  Tsai  is  a  graduate  student  in  the  Dept  of  Civil  and 
Environmental  Engineering.  T.  Steenhuis  is  an  associate  professor  in  the  Dept 
of  Agricultural  and  Biological  Engineering.  C.McCulloch  is  a  professor  in  the 
Biometrics  Unit  and  Statistics  Center. 

REFERENCES 

Bartell,  S.M.  and  A.L.  Brenkert.  1991.  A  Spatial-temporal  model  of  nitrogen 
dynamics  in  a  deciduous  forest  watershed.,  pp  379-398  in:  Turner,  M.  G.  and 
R.H.  Gardner  (eds.)  1991.  Quantitative  Methods  in  Landscape  Ecology 
Ecological  Studies  82,  Springer- Verlag,  NY. 

Beven,  K.J.  and  M.J.  Kirkby.  1979.  A  physically-based  variable  contributing 
area  model  of  basin  hydrology.  Hydrol.  Sci.  J.  ,  24(l):43-69 

Johnsson,  H.  1990.  Nitrogen  and  water  dynamics  in  arable  soil.  Swedish 
University  of  Agricultural  Sciences,  Department  of  Soil  Sciences  Reports  and 
Dissertations,  6.  Uppsala,  Sweden. 

Kuo,  W-L.,  P.  Longabucco,  M.R.  Rafferty,  J.  Boll,  and  T.S.  Steenhuis.  1996. 
An  integrated  GIS -based  model  for  soil  water  and  nitrogen  dynamics  in  a 
New  York  City  watershed.  Proc.  AWRA  Symposium  "Watershed 
Restoration  Management"  (July  14-17,  1996).  Syracuse,  New  York 
(Accepted) 

McCulloch,  C.E.,  W.L.  Kuo,  and  D.P.  Swaney.  1995.  Designer  landscapes: 
Simulating  landscapes  for  management  of  change.  Invited  Presentation, 
Eastern  North  American  Region  Biometrics  Meeting.  Birmingham,  AL 

Poiani,  K.  A.,  B.L.  Bedford  and  M.D.  Merrill.  1996.  A  GIS-based  index  for 
relating  landscape  characteristics  to  potential  nitrogen  leaching  in  wetlands. 
Landscape  Ecology  (submitted). 

Zollweg,  J.  A.  1994.  Effective  use  of  geographic  information  systems  for 
rainfall-runoff  modeling.  Unpublished  dissertation,  Cornell  University. 


72 


Understanding  the  Spatial 
Distribution  of  Tree  Species  in 
Pennsylvania 

Rachel  Riemann  Hershey1 


Abstract.— Current,  accessible  information  on  the  distribution  of  tree 
species  would  aid  in  the  understanding  and  management  of  ecosystems. 
However,  such  detailed  information  on  forest  composition  is  only 
available  from  ground  inventory.  Geostatistical  techniques  are  used  here 
to  create  an  interpolated  dataset,  a  'map'  of  individual  species 
distribution,  from  known  sample  information.  In  a  previous  study,  we 
found  that  indicator  kriging  and  sequential  gaussian  conditional 
simulation  (sgCS)  were  promising  tools  for  estimating  sugar  maple 
distribution  from  the  USDA  National  Forest  Inventory  and  Analysis 
(FIA)  data.  The  techniques  provided  an  estimate  of  species  occurrence 
and  a  measure  of  uncertainty  associated  with  that  estimate,  while 
retaining  much  of  the  local  variability  present  in  the  sample  data.  In  this 
study,  these  techniques  are  applied  to  9  additional  species  in 
Pennsylvania.  Four  output  datasets  are  available  for  each  species—the 
probability  of  species  occurrence,  an  estimate  of  its  relative  abundance, 
and  a  plus  and  minus  level  of  uncertainty  associated  with  that  estimate. 
The  datasets,  used  in  conjunction  with  one  another,  provide  the  user  with 
considerable  flexibility  in  setting  up  the  balance  of  errors  of  omission 
and  commission  that  best  suit  the  analysis  under  consideration. 
Similarities  and  differences  between  the  species  are  identified  and 
discussed  as  to  their  possible  effect  on  the  final  estimates.  Examples  of 
how  the  datasets  can  be  used  are  also  presented.  Indicator  kriging  and 
sgCS,  used  in  conjunction  with  FIA  sample  data,  provide  a  relatively 
straightforward  technique  to  describe  species  occurrence  and  relative 
density  across  a  state. 


INTRODUCTION 

Data  describing  forest  composition— so  desired  as  a  basic  data  source  for  many 
aspects  of  ecosystem  analyses,  models,  and  management— are  generally 
unavailable  and/or  is  stored  in  fixed  forest-type  categories.  But  forest 
communities  are  often  not  well  characterized  by  the  discrete  categories  imposed 
by  forest  cover  type  divisions.  Inherent  in  each  forest  type  category  is  an  entire 
continuum  (usually  multi-dimensional)  of  different  species  and  their  relative 
importance. 

In  a  previous  study,  we  compared  the  geostatistical  techniques  available  to 
interpolate  FIA  sample  data  to  create  a  'map'  of  tree  species  distribution.2  The 


Northeastern  Forest  Experiment  Station.  USDA  Forest  Service,  5  Radnor  Corporate  Center.  Radnor.  PA  19087-4585. 


73 


tools  of  ordinary  kriging,  multigaussian  kriging,  indicator  kriging,  and  sequential 
gaussian  conditional  simulation  (sgCS)  were  used  to  estimate  the  occurrence  and 
distribution  of  sugar  maple  in  Pennsylvania.  After  considering  the  phenomenon 
being  examined,  the  sample  data  being  used,  and  the  kind(s)  of  output  desired  in 
this  study,  we  decided  that  indicator  kriging  sgCS  proved  to  be  the  best 
interpolation  tools  for: 

a)  providing  an  estimate  of  sugar  maple  occurrence, 

b)  providing  an  estimate  of  sugar  maple  'importance'  in  terms  of  %ba/acre, 

c)  providing  a  measure  of  uncertainty  associated  with  the  two  estimates, 

d)  maintaining  local  variability, 

e)  maintaining  the  characteristics  of  the  original  sample  data,  and 

f)  handling  sample  data  with  highly  skewed  distributions. 

a)  An  estimate  of  each  species'  presence  or  absence  was  provided  by  indicator 
kriging.  An  indicator  transform  divides  the  data  into  two  classes—either  above  or 
below  a  designated  cutoff  value;  in  this  study  0%ba/acre  indicating  presence  or 
absence.  Indicator  kriging  calculates  for  each  cell  an  estimate  of  the  probability 
that  it  falls  above  or  below  the  cutoff  value.  The  output  dataset  thus  indicated  the 
probability  that  sugar  maple  occurred  at  each  location. 

b)  The  second  piece  of  information  desired  was  an  estimate  of  the  relative 
amount  of  sugar  maple  at  that  location— i.e.,  whether  the  species  represented  a 
minor,  moderate,  or  a  major  component  of  the  total  ba/acre  on  the  plot  at  that 
location.  Sequential  gaussian  conditional  simulation  determines  multiple 
estimates  for  each  cell.  All  are  equally  probable,  and  yet  alternative  realizations 
of  the  data  determined  from  multiple  simulation  runs.  From  this  set  of  estimates, 
an  entire  distribution  can  be  built  for  each  cell,  representing  the  range  of  possible 
values.  A  summary  statistic  such  as  the  mean  or  median  of  this  distribution  can 
then  be  chosen  and  used  as  the  modeled  'estimate'  of  %ba/acre  for  that  cell.3 

c)  A  level  of  uncertainty  is  always  associated  with  any  estimate.  Knowing  how 
much  uncertainty  exists  will  help  the  user  identify  whether  that  uncertainty  is 
acceptable  for  a  specific  task  and  how  the  data  can  be  used.  Correspondingly, 
knowing  how  much  uncertainty  exists  will  help  the  producer  identify  areas  in 
which  additional  sampling  would  most  improve  the  estimates.  For  estimates  of 
species'  presence/absence,  indicator  kriging  provides  this  information  in  terms  of 
a  probability.  For  estimates  of  %ba/acre,  summary  statistics  such  as  standard 
deviation  or  inter-quartile  range  were  calculated  from  the  distribution  of  simulated 
values  to  describe  the  variation  associated  with  the  %ba/acre  estimate  for  each 
cell. 


^  Hershey,  R.  Riemann,  M.A.  Ramirez,  and  D.A.  Drake.  Using  Geostatistical  techniques  to  estimate  the  distribution  and 
relative  density  of  individual  tree  species  in  Pennsylvania.  Unpublished  report  on  file  at  USDA  Forest  Service, 
Northeastern  Forest  Experiment  Station,  Forest  Inventory  and  Analysis  Unit,  Radnor,  PA. 

3  The  methods  are  described  in  Rossi  et  al.  1993  and  Isaaks  and  Srivastava  1989;  the  analysis  was  performed  using 
GSLIB  routines  (Deutsch  and  Journel  1992)  with  some  additional  routines  written  by  R.E.  Rossi. 


74 


d)  Tree  species  in  Pennsylvania  exhibit  a  high  level  of  local  variation  as  a  result 
of  natural  environmental  factors  and  land  use  histories.  At  the  intensity  of 
sampling  present  in  the  FIA  sample  data,  much  of  this  local  variation  cannot  be 
modeled  and  effectively  predicted  in  the  interpolation  process,  but  instead  appears 
as  variation  that  is  unexplained  by  neighboring  plots.  However,  such  local 
variability  is  an  important  characteristic  of  the  distribution  of  a  species.  Thus,  we 
did  not  want  this  local  variability  to  become  hidden  behind  a  regional  average  of 
the  resource,  but  to  remain  as  apparent  and  accessible  to  the  user  as  possible  in  the 
final  estimated  dataset(s).  Sequential  gaussian  conditional  simulation  was  the 
most  effective  of  the  interpolation  methods  at  maintaining  local  variation. 

e)  One  feature  of  a  well-designed  sampling  scheme  is  that  it  is  sensitive  to  and 
can  report,  with  an  acceptable  level  of  error,  the  characteristics  of  the  phenomena 
of  interest.  In  this  ideal  situation,  the  characteristics  of  the  sample  data  represent 
reasonably  well  those  characteristics  of  the  phenomena  itself.  Every  estimation 
technique  honors  and  maintains  different  aspects  of  the  original  data.  The  specific 
goals  of  the  interpolation  task  at  hand  will  determine  the  priorities,  but  in  general 
the  more  characteristics  of  the  sample  data  that  are  preserved  in  the  estimated 
dataset,  the  more  desirable  the  dataset.  Sequential  gaussian  conditional  simulation 
again  did  the  best  job  of  maintaining  both  the  univariate  and  bivariate 
characteristics  of  the  sample  data. 

f)  As  is  true  with  many  plant  and  animal  populations,  tree  species  have 
population  distributions  that  are  distinctly  skewed  toward  younger  individuals- 
more  small  trees  than  large  mature  ones.  In  addition,  Pennsylvania,  like  most  of 
the  northeastern  states,  contains  primarily  mixed  forests.  Individual  species  rarely 
occur  in  pure  stands.  The  10  species  examined  in  both  studies  included  8  of  the 
most  common  by  volume  in  Pennsylvania,  and  yet  more  than  50%  of  the  time, 
when  a  species  occurred  on  a  plot  it  occurred  as  only  a  minor  component  (here 
defined  as  making  up  less  than  20%  of  the  total  ba/acre  on  that  plot).  Both  factors 
are  combined  in  the  %ba/acre  'relative  importance'  value,  resulting  in  a  highly 
skewed  frequency  distribution.  Such  extreme  characteristics  in  the  sample  data 
can  cause  difficulties  and  biases  when  used  with  some  of  the  interpolation 
methods  that  depend  on  assumptions  about  the  normality  of  the  distribution  of  the 
sample  data  (Isaaks  and  Srivastava  1989).  One  particular  advantage  of  indicator 
kriging  is  that  it  makes  no  assumptions  about  the  distribution  of  the  data. 
Sequential  gaussian  conditional  simulation,  on  the  other  hand,  does  assume  that 
the  data  are  normally  distributed  and  stationary,  and  must  be  used  more  carefully. 
The  sgCS  routine  used  here,  from  Deutsch  and  Journel  (1992),  performs  a  1-1, 
invertible  normal-score  transform  on  the  data  before  running  the  simulation.  In 
addition,  however,  the  data  also  should  be  checked  for  binormality  and  a  decision 
made  as  to  whether  to  assume  multivariate  normality  before  the  results  of 
conditional  simulation  are  accepted  (Rossi  et  al.  1993). 

The  data  exploration  techniques  used  also  proved  invaluable  to  understanding 
the  spatial  characteristics  of  the  species/variable  being  examined.  Techniques 
included  univariate  analysis,  variograms  and  other  spatial  dependence  analyses, 


75 


and  calculating  local  statistics.  The  resulting  information  was  critical  not  only  in 
determining  what  interpolation  methods  were  most  suitable  and  for  checking  the 
sample  data  for  errors,  but  also  for  understanding  the  characteristics  of  the  sample 
data  and  thus  the  phenomena  being  investigated.  Geostatistical  techniques 
offered  ways  to  explore,  organize,  and  summarize  spatial  patterns  in  the  data  that 
can  provide  clues  to  the  variation  and  spatial  behavior  of  the  individual  species 
under  investigation. 

Applying  the  techniques  to  10  species  in  Pennsylvania 

Because  of  the  promising  results  from  using  geostatistical  techniques  for 
estimating  the  distribution  of  sugar  maple  from  FIA  data  in  the  previous  study,  the 
same  geostatistical  methods  were  applied  to  nine  additional  species:  red  oak, 
white  oak,  chestnut  oak,  black  oak,  hemlock,  red  maple,  beech,  white  pine,  and 
yellow  birch.  This  list  includes  8  of  the  top  10  most  abundant  species  in 
Pennsylvania  by  volume,  and  two  species  (yellow  birch  and  white  pine)  that  are 
much  less  common  (Alerich  1993). 

Tree  species  distribution  is  affected  by  many  factors,  including  both 
environmental  conditions  and  direct  human  influence  through  harvesting  and 
other  land  use  histories.  As  a  result  of  being  differentially  affected  by  all  of  these 
factors,  each  species  will  exhibit  different  patterns  and  scales  of  spatial 
distribution.  Some  of  these  factors  occur  at  scales  much  smaller  than  the  sampling 
intensity  of  the  FIA  data,  and  some  occur  over  larger  areas,  representing  broad- 
scale  variation  in  the  species  distribution.  In  the  previous  study,  it  was  found  that 
a  substantial  amount  of  variation  in  sugar  maple  distribution  was  resolved  at  the 
sampling  scale  used  for  the  FIA  plots.  This  spatial  dependence  could,  therefore, 
be  modeled  and  used  to  support  estimates  of  species  occurrence  and  relative 
'importance'  (%ba/acre).  The  goal  of  this  study  is  to  examine  to  what  extent  this 
is  true  for  the  other  species.  More  specifically,  the  objectives  of  this  study  are: 

a)  if  spatial  dependence  is  exhibited  at  this  intensity/scale  of  sampling, 

b)  the  resulting  spatial  distribution  for  each  species,  and  how  that  compares  to 
our  current  understanding, 

c)  how  the  species  differ  from  one  another  in  terms  of  spatial  dependence  and 
distribution,  and  how  that  affects  our  ability  to  estimate  them,  and 

d)  how  to  use  the  resulting  estimated  datasets. 

DATA 

The  sample  data  were  collected  by  the  Northeastern  Forest  Experiment  Station's 
Forest  Inventory  and  Analysis  (FIA)  unit.  Basal  area—the  summed  cross-sectional 
area  at  breast  height-is  calculated  for  all  live  trees  1 .0  inches  DBH  or  larger  on 
the  plot  (Hansen  et  al.  1992).  The  data  were  for  individual  tree  species,  by  basal 
area  (ba)  per  acre  as  a  proportion  of  the  total  basal  area  (%  ba/acre).  The  data 
were  accessed  from  individual  tree  records  in  the  USFS  Eastwide  tree-level 
database  and  summarized  as  %ba/acre  for  each  species  by  plot.  In  Pennsylvania, 
there  were  a  total  of  5,100  plots.  Nonforested  plots  and  those  with  total  ba/acre 
equal  to  zero  (due  to  missing  data)  were  removed-leaving  only  2,905  plots. 


76 


METHODS 


Each  species  was  examined  entirely  independently.  As  in  the  previous  study, 
the  data  for  each  species  was  organized,  summarized,  and  explored  using 
univariate  statistics,  measures  of  spatial  dependence  (variogram,  covariance,  and 
correlogram),  and  spatial  distribution  of  local  statistics  across  the  state.  All 
species  were  similar  in  many  of  their  basic  characteristics  to  each  other  and  to  the 
previously  investigated  species,  sugar  maple.  Each  exhibited  extremely  skewed 
distributions,  with  more  than  50%  of  the  plots  containing  less  than  l%ba/acre  in 
every  species  except  red  maple  and  red  oak.  A  variogram  was  calculated  for  both 
the  raw  sample  data  and  for  a  1-1,  invertible  normal-score  transform  of  the  sample 
data,  using  a  lag  distance  of  500m  and  no  directional  component  (anisotropy).  In 
every  instance,  the  variogram  of  the  normal-scored  data  exhibited  considerably 
more  spatial  dependence  and  structure  than  that  of  the  raw  data  (Figure  1), 
revealing  spatial  characteristics  that  were  hidden  by  the  strong  univariate 
characteristics  of  the  data.  As  sgCS  uses  normal-scored  data,  it  was  the  model 
fitted  to  the  normal-scored  variogram  that  was  used  in  the  conditional  simulation. 
An  indicator  variogram  also  was  calculated  and  modeled  for  use  in  the  indicator 
kriging.  In  general,  there  was  far  less  structure  and  less  of  the  variation  explained 
in  the  indicator  variogram  (32  to  57%)  than  in  the  normal-scored  regular 
variogram  (35%  and  64  to  97%)  (Table  1).  To  assess  how  areas  of  'local' 
variability  in  the  sample  data  changed  across  the  state,  the  mean  and  standard 
deviation  were  calculated  for  each  of  the  23,400  3000  x  3000m  cells,  using  a  15  x 
15km  area  as  the  window  defining  the  size  of  the  'local'  area.  All  species 
exhibited  a  proportional  effect,  with  areas  of  high  mean  corresponding  with  areas 
of  high  local  standard  deviation,  indicating  a  lack  of  stationarity.  Using  normal- 
scored  data  seemed  to  largely  eliminate  this  situation. 

Table  1.  The  percent  variation  explained  by  the  spatial  dependence  in  the  variograms. 


Indicator 

Normal-scores 

Species 

variogram 

variogram 

Beech 

46 

76 

Black  oak 

37 

75 

Chestnut  oak 

44 

82 

Hemlock 

57 

66 

Red  maple 

32 

35 

Red  oak 

38 

64 

Sugar  maple 

32 

97 

White  oak 

50 

79 

White  pine 

37 

84 

Yellow  birch 

39 

81 

Indicator  kriging  and  sgCS  were  run  for  each  species,  using  models  derived 
from  the  appropriate  variograms.  The  estimation  parameters  of  cell  size  (3,000m), 
search  radius  (10,000m),  and  minimum:maximum  number  of  points  used  (1:16) 
were  taken  directly  from  the  results  of  the  previous  study.  In  sgCS,  50 
simulations  were  run  for  each  species. 


77 


0        25000    50000    75000  100000 
distance  (m) 


0        25000     50000     75000  100000 
distance  (m) 


Figure  1.  Spatial  dependence  as  demonstrated  by  the  variogram  of  the  raw  data  (left  side) 
and  variogram  of  the  normal-scored  data  (right  side)  for  white  pine. 

RESULTS 

With  the  exception  of  red  maple,  all  species  examined  demonstrated  substantial 
spatial  dependence  in  the  variogram  of  normal-scored  data,  with  64  to  97%  of  the 
variation  explained  by  the  visible  structure  and  capable  of  being  modeled.  In 
some  species,  much  of  that  spatial  dependence  was  contained  in  a  very  long-range 
trend  of  about  100,000m.  White  oak,  black  oak,  chestnut  oak,  and  beech  all  fell 
into  this  category.  The  rest  of  the  species  appeared  to  split  the  bulk  of  the 
explained  spatial  dependence  over  2  ranges.  For  red  oak,  this  was  both  a  short- 
and  medium-range  pattern  (12,000  and  40,000m);  sugar  maple  a  very  short-  and  a 
long-range  pattern  (2,100  and  60,000m);  and  yellow  birch  a  medium-  and  long- 
range  pattern  (19,000  and  80,000)  (Figure  2).  The  spatial  dependence  exhibited 
in  the  indicator  variograms  was  much  less,  ranging  from  only  32  to  57%  of  the 
variation  explained.  This  is  not  ideal  for  interpolation  and  suggests  the  necessity 
for  further  refinement. 


distance  (m)  distance  (m) 

^sample  data  —model 


0         25000      50000      75000     100000  0        25000      50000      75000  100000 


distance  (m)  distance  (m) 

Figure  2a-d.  Variograms  from  the  normal-scored  data  for  four  of  the  species. 

The  results  create  several  output  datasets.  Figure  3a-d  shows  the  results  for 
beech,  using  4  datasets  to  represent  the  species.  Part  (a)  shows  the  estimated 
probability  of  beech  occurrence,  as  calculated  from  indicator  kriging.  Part  (b)  is 
the  median  value  of  the  50  sgCS  simulations,  representing  the  chosen  estimate  of 
beech  relative  'importance'  in  %ba/acre.  The  uncertainty  associated  with  that 
estimate  (here  chosen  to  be  a  percentile  range  capturing  approximately  2/3  of  the 


78 


distribution)  is  described  in  (c)  and  (d).  Part  (c)  expresses  the  minus  variation,  or 
that  distance  in  %ba/acre  values  between  the  median  and  the  bottom  of  that  range 
(the  17th  percentile),  and  part  (d)  expresses  the  plus  (+)  variation  (83rd-50th 
percentiles).  In  every  instance,  the  +  variation  is  much  greater. 


Figure  3.  Four  estimated  datasets  describing  the  distribution  of  beech  in  Pennsylvania:  a)  the 
estimated  probability  of  occurrence  using  indicator  kriging,  b)  the  median  values  from  50 
sequential  gaussian  conditional  simulations,  c)  the  minus  variation  (median-1 7th  percentile), 
and  d)  the  plus  variation  (83rd  percentile-median)  about  the  median  estimate. 


DISCUSSION 


The  probability  that  a  species  occurs  is  a  unique  and  useful  dataset.  Any 
probability  can  be  used  as  the  cutoff  to  create  a  species  presence/absence  map 
depending  on  the  objectives  of  the  task  at  hand.  If,  for  example,  a  particular 
insect  is  known  to  live  in  chestnut  oak  forests  and  the  objective  is  to  limit  the 
search  to  only  those  areas  where  there  is  a  high  probability  of  finding  suitable 
conditions,  we  might  set  the  cutoff  at  the  probability  level  of  >  .8.  If,  however, 
we  are  most  interested  in  not  missing  any  areas  where  the  insect  occurred,  we 
might  set  the  probability  level  for  forest  much  lower,  say  >  .4. 


Figure  4.  Chestnut  oak  occurrence  as  a  Figure  5.  The  uncertainty  associated  with 

major  (>40%ba/acre),  moderate  (20-40%),  the  estimates  used  in  Figure  4,  in  classes  of: 

and  minor  component  (<20%)  of  the  total  acceptable  (<±10%ba/acre),  moderate  (±10- 

ba/acre.  Derived  from  sgCS  estimates  using  40%),  and  unacceptable  (>±40%).  Derived 

the  75th  percentile.  from  the  sgCS  estimates. 


79 


The  distribution  of  densities  at  which  a  species  occurs  also  can  be  of 
considerable  interest.  For  example,  Figure  4  is  a  plot  produced  from  sgCS 
estimates  illustrating  where  chestnut  oak  occurs  as  a  major,  moderate,  and  minor 
component.  The  three  categories  have  been  defined  as  1  to  20%ba/acre,  20  to 
40%ba/acre,  and  >40%ba/acre.  The  %ba/acre  estimate  in  Figure  4  can  be 
associated  with  a  corresponding  uncertainty  dataset  (Figure  5)  in  which  the 
uncertainty  classes  are  broken  into  acceptable,  moderate,  and  unacceptable. 
Again,  the  dataset  itself  is  unclassed,  and  the  user  defines  these—all  the  building 
blocks  are  provided  in  the  output  from  sgCS. 

Alternatively,  the  same  %ba/acre  dataset(s)  can  be  used  to  create  a  forest  cover 
type  'map.'  For  this  purpose,  the  summary  statistic  (whether  mean,  median,  or 
another  of  the  percentiles  of  each  cell's  simulated  distribution)  can  be  chosen 
specifically  for  the  intended  purpose  in  the  same  way  the  different  levels  of 
probability  could  be  chosen  when  mapping  species  presence/absence  from 
indicator  kriging  estimates.  If  the  objective  is  to  reduce  the  error  of  commission 
(i.e.,  classifying  areas  as  sugar  maple/beech  (SM/B)  in  the  estimated  map  that 
really  are  not  SM/B),  then  using  a  percentile  at  the  lower  end  of  each  cell's 
distribution  would  be  more  desirable.  If,  however,  the  objective  is  to  reduce  the 
error  of  omission  (i.e.,  missing  areas  that  do  contain  SM/B),  then  using  a  higher 
percentile  from  the  distribution,  such  as  75%,  would  be  more  desirable. 

The  %ba/acre  estimates  of  three  species  were  further  refined.  Two  species, 
white  pine  and  yellow  birch,  nearly  disappeared  in  the  initial  sgCS  estimate  even 
though  they  showed  up  as  present  in  the  probability  of  occurrence  estimate  from 
the  indicator  kriging.  This  was  considered  suspicious,  so  the  dataset  for  each  of 
these  two  species  was  divided  into  several  different  populations  by  region,  and  the 
process  of  variogram  modeling  and  interpolation  was  repeated  for  each  region. 
The  regionalized  variogram  was  substantially  different  from  the  average 
variogram,  and  when  sgCS  was  performed  using  the  locally  tuned  models,  it 
revealed  an  entirely  different  and  much  more  credible  picture  of  white  pine 
distribution  in  Pennsylvania.  The  third  species,  hemlock,  was  refined  based 
primarily  on  the  suspicion  that  small  and  large  stand-size  classes  may  have 
different  spatial  distribution  patterns.  When  variograms  were  calculated 
separately  for  these  two  populations  (using  a  cutoff  of  45  years),  they  were  indeed 
substantially  different  in  shape,  sill,  nugget,  and  range. 

CONCLUSIONS 

The  estimated  datasets  output  by  indicator  kriging  and  sgCS  have  the  potential 
to  be  very  useful.  Every  species  examined,  with  the  exception  of  red  maple, 
exhibited  substantial  spatial  dependence  in  the  variograms  of  the  normal-scored 
data,  suggesting  that  there  is  considerable  potential  for  the  estimation  of  such 
datasets  from  FIA  data.  Each  species  exhibited  some  variety  in  spatial  patterns 
and  spatial  dependence  and  may  require  different  levels  of  additional  fine-tuning, 
depending  upon  the  objectives  of  the  specific  analysis  and  the  time  and  expertise 
available. 


80 


Characteristics  that  make  the  datasets  useful 

These  techniques  make  explicit  the  uncertainties  associated  with  an  estimate  in 
a  form  that  can  be  incorporated  when  the  data  are  used.  This  feature  adds 
considerable  utility  and  flexibility  in  the  use  of  the  resulting  estimates,  as  the  risk 
of  errors  of  commission  or  omission  can  be  specifically  determined  and 
manipulated  to  suit  the  current  objectives. 

Maintaining  individual  species  information  separately  allows  considerable 
flexibility  in  the  use  of  species  distribution  data.  Instead  of  being  limited  to 
previously  defined  fixed  classes,  forest  cover  types  can  be  uniquely  defined  to 
capture  more  accurately  the  habitat  required  for  a  particular  study.  The  potential 
also  exists  to  use  one  or  more  of  the  species  datasets  as  a  decision  layer  in  the 
interpretation  of  satellite  imagery.  The  two  datasets  offer  complementary 
information  about  the  species  composition  that  really  exists  on  the  ground. 

The  techniques  used  in  this  study  are  not  extremely  time-consuming  nor 
difficult  to  process,  and  could  be  easily  extended  to  additional  species  and  states. 
There  is  a  high  level  of  variance  associated  with  these  estimates  of  %ba/acre— in 
many  locations  this  variance  can  be  as  much  as  the  estimate  itself.  However,  the 
dataset  nevertheless  provides  a  very  descriptive  picture  of  species  distribution  at 
the  state  level.  In  comparison  to  previous  depictions  of  current  species 
distribution  from  FIA  data  by  summarizing  at  the  county  level,  this  method 
provides  a  much  more  detailed  picture  of  species  occurrence  and  distribution. 
Although  estimates  could  probably  be  improved  and  variances  diminished  by 
additional  investigation  into  each  species,  the  current  estimates  are  informative 
and  provide  a  useful  basis  from  which  to  proceed. 

Clues  to  refining  the  interpolation  and  improving  the  estimates 

As  was  observed  with  white  pine,  yellow  birch,  and  hemlock,  it  may  be  possible 
to  significantly  improve  the  estimates  of  %ba/acre  by  refining  the  analysis.  When 
sub  populations  of  a  species  have  a  significantly  different  pattern  of  spatial 
distribution,  treating  the  populations  separately  in  the  interpolation  will  improve 
the  final  estimates.  These  populations  may  be  described  by  regional  land  features 
or  by  some  other  defining  characteristic  (e.g.,  stand  age  for  hemlock).  The  results 
suggest  several  clues  to  determine  when  this  additional  effort  is  necessary.  First, 
when  results  of  indicator  kriging  showed  a  species  as  present,  but  sgCS  did  not. 
Second,  when  the  spatial  dependence  in  the  variogram  is  noticeably  less  than 
expected.  This  may  suggest  that  there  are  different  populations  being  lumped 
together  that  should  be  separated.  For  white  pine,  dividing  the  state  into  several 
broad  geographic  regions  made  a  significant  difference  in  the  calculated 
variogram  and  thus  in  the  final  estimates.  Another  important  clue  is  previous 
knowledge  about  the  species  that  different  ecological  regions  may  have  caused 
distinct  spatial  distribution  patterns  among  tree  species,  or  that  different  size  class 
or  age  populations  may  have  different  spatial  distribution  patterns  over  the 
landscape.  Hemlock  is  an  example  of  the  latter.  As  a  result  of  past  management 
practices  that  involved  heavy  harvesting  of  large  hemlock  for  the  tanning 
industry,  today  there  are  often  relics  of  large  individuals  among  a  relatively  wider 


81 


distribution  of  smaller,  younger  trees  that  have  grown  up  in  the  interim  (Hough 
and  Forbes  1943,  Powell  and  Considine  1982). 

These  datasets  of  individual  species  distribution  do  not  contain  any  of  the  fine- 
scale  forest/nonforest  detail.  If  such  information  is  desired,  more  detailed  datasets 
describing  the  forest/nonforest  land  cover  in  Pennsylvania  would  have  to  be 
derived  from  a  more  intense  point  sample  or  the  continuous  but  averaged  data 
available  from  satellite  imagery.  Such  detailed  datasets  could  be  used  as  a  'mask' 
and  overlaid  on  any  of  the  datasets  of  species  distribution. 

There  are  more  possibilities  for  applying  geostatistical  techniques  than  have 
been  investigated  here.  For  example,  some  species  may  exhibit  some  correlation 
with  a  particular  soil,  climate,  topography,  or  reflectance  data  from  satellite 
imagery.  Indicator  kriging  and  sgCS,  in  particular,  allow  the  incorporation  of  such 
ancillary,  'soft'  information  to  contribute  to  the  estimation  process. 

REFERENCES 

Alerich,  C.L.  1993.  Forest  Statistics  for  Pennsylvania— 1978  and  1989.  Resource 
Bulletin  NE-126.  USDA  Forest  Service,  Northeastern  Forest  Experiment 
Station.  Radnor,  PA.  244p. 

Deutsch,  C.V.  and  A.G.  Journel.  1992.  GSLIB:  Geostatistical  Software  Library 
and  User's  Guide.  Oxford  University  Press,  New  York. 

Hansen  M.H.,  T.  Frieswyk,  J.F.  Glover,  and  J.F.  Kelly.  1992.  The  Eastwide  Forest 
Inventory  Data  Base:  Users  Manual.  General  Technical  Report  NC-151.  USDA 
Forest  Service,  North  Central  Experiment  Station.  St.  Paul,  MN. 

Hough,  A.F.  and  R.D.  Forbes.  1943.  The  Ecology  and  silvics  of  forests  in  the 
High  Plateaus  of  Pennsylvania.  Ecological  Monographs.  13:299-320. 

Isaaks,  E.H.  and  R.M.  Srivastava.  1989.  An  Introduction  to  Applied  Geostatistics. 
Oxford  University  Press,  New  York. 

Powell,  D.S.  and  T.J.  Considine.  1982.  An  analysis  of  Pennsylvania's  forest 
resources.  Resource  Bulletin  NE-69.  USDA  Forest  Service,  Northeastern  Forest 
Experiment  Station.  Broomall,  PA.  97p. 

Rossi,  R.E.,  P.W.  Borth,  and  J.J.  Tollefson.  1993.  Stochastic  simulation  for 
characterizing  ecological  spatial  patterns  and  appraising  risk.  Ecological 
Applications.  3(4):719-735. 

BIOGRAPHICAL  SKETCH 

Rachel  Riemann  Hershey  is  a  forester/geographer  with  the  Forest  Inventory  and 
Analysis  Unit,  Northeastern  Forest  Experiment  Station.  She  received  a  B.A.  in 
ecology  from  Middlebury  College,  an  M.S.  in  forestry  from  the  University  of  NH 
and  an  M.Phil,  in  geography  from  the  London  School  of  Economics. 


82 


A  Calibration-Based  Model  for 
Correcting  Area  Estimates  From 
Coarse  Resolution  Land  Cover  Data 

Aaron  Moody 


Abstract. — A  two  stage  modeling  strategy  significantly  improves  area 
estimates  by  correcting  coarse-resolution  measurements  of  class  propor- 
tions. Stage  I  models  use  measurements  of  scale-invariant  landscape  spa- 
tial properties  to  estimate  the  slope  and  intercept  of  proportion  transition 
relationships.  A  stage  II  model  uses  a  regression  estimator  to  predict  true 
class  proportions  based  on  measured  coarse-scale  proportions,  and  the 
slope  and  intercept  estimates  from  the  stage  I  models.  Model  develop- 
ment and  testing  on  a  calibration  site  is  followed  by  testing  and  inversion 
for  a  validation  site.  Inversion  involves  using  spatial  variables  measured 
at  the  coarse  scale  as  input  to  the  stage  I  models.  A  probabilistic  sam- 
pling strategy  allows  statistical  assessment  of  the  models  and  results. 


INTRODUCTION 

Image  spatial  resolution  will  influence  land-cover  area  estimates  made  from 
classified  remotely  sensed  data  (Mayaux  and  Lambin  1995;  Moody  and  Wood- 
cock 1994).  Models  of  these  resolution  effects  can  lead  to  improved  area  esti- 
mates derived  from  coarse  resolution  remote  sensing.  One  might  consider  two 
approaches  for  modeling  this  scale-dependent  areal  bias.  Mixture  models,  incor- 
porated into  the  classification  process,  can  estimate  the  subpixel  composition  of 
pixels  if  pure-class  spectra  are  known  (Adams  et  al.  1986).  Alternatively,  cali- 
brated models,  applied  in  a  post-classification  mode,  can  improve  area  estimates 
from  coarse-resolution  data  if  the  relationship  between  "true"  and  measured  pro- 
portions can  be  modeled  (Mayaux  and  Lambin  1995;  Kalkhan  et  al.  1995). 

This  paper  describes  the  development  and  evaluation  of  three  linked,  statisti- 
cal models  that  provide  post-classification  correction  of  area  estimates.  The  first 
two  models  estimate  the  slope  and  intercept  of  a  line  characterizing  the  relation- 
ship between  true  and  coarse-scale  proportions.  In  this  case,  "true"  proportions 
refer  to  proportions  measured  at  30  m,  and  coarse-scale  proportions  refer  to  meas- 
urements at  1020  m.  The  slope  and  intercept  are  modeled  based  on  a  small  set  of 
relatively  scale-invariant,  measurable  spatial  properties  of  the  landscape.  The 
third  model  uses  these  estimated  slopes  and  intercepts  to  predict  30  m  proportions 
based  on  measurements  of  proportions  at  1020  m. 


Assistant  Professor,  Department  of  Geography,  University  of  North  Carolina,  Chapel  Hill,  NC.  27599 


83 


BACKGROUND 


Two  basic  spatial  effects  contribute  to  biased  area  estimates.  A  first  order 
effect  is  the  tendency  of  large  classes  to  increasingly  dominate  the  landscape 
when  measured  at  increasingly  coarse  scales.  Accordingly,  small  classes  tend  to 
diminish  in  size.  Figure  1  illustrates  these  patterns  for  all  the  data  used  in  this 
study.  Second  order  effects  refer  to  modulations  of  these  basic  patterns  due  to 
specific  landscape  spatial  organization  (Moody  and  Woodcock  1995).  These 
effects  result  in  the  scatter  about  the  smoothed  fit  notable  in  Figure  Id.  In  either 
case,  scale-dependent  changes  in  the  apparent  area  of  classes  result  from  class 
membership  transitions  between  scales.  This  effect  can  be  thus  characterized  in 
terms  of  proportion  transition  lines  relating  true  and  coarse  resolution  proportions. 
The  slope  of  such  a  line  can  summarize  the  rate  of  transition. 

If  transition  rates  depend  partly  on  landscape  spatial  organization,  it  is  sensi- 
ble to  try  to  model  them  using  measures  of  spatial  pattern.  A  variety  of  spatial 
measures  exist,  and  many  reviews  and  summaries  are  present  in  vegetation 
analysis  and  landscape  ecology  literature  (Legendre  and  Fortin  1989;  Cullinan 
and  Thomas  1992).  In  this  research,  landscape  spatial  properties  are  used  to 
model  the  slopes  and  intercepts  of  the  proportion  transition  lines  for  a  large  set  of 
sampling  units  in  independent  calibration  and  validation  sites.  These  are  the 
stage  I  models.  Model  inversion,  will  require  that  the  landscape  pattern  measures 
are  relatively  scale-invariant,  or  resistant  to  resolution. 


Figure  1. — Relationship  between  proportions  at  30  m  and  at  4  coarser  scales. 


84 


There  are  several  calibration-based  methods  for  improving  coarse  resolution 
area  estimates  (Czaplewski  and  Catts  1992;  Kalkhan  et  al.  1995).  The  model 
used  here  is  a  form  of  "classical"  model  in  which  known  but  incorrect  values  are 
estimated  using  unknown  correct  values  (Brown  1982).  For  example,  if  Pq 
represents  true  proportions  and  Pr  represents  measured  proportions  at  some 
coarse  spatial  resolution  r ,  then: 

Pr  =p0  +  pr/,o  + error  (1) 

where  po  pi  are  the  intercept  and  slope  of  the  proportion  transition  line  that 
relates  true  proportions  to  proportions  measured  at  resolution  r .  Inverting  this 
simple  model  provides: 

Pr  ~  (3f) 

which  is  the  stage  II  model  used  in  this  paper. 

A  five  part  strategy  is  employed  as  follows:  a)  identify  a  set  of  scale  invari- 
ant spatial  measures;  b)  use  a  subset  of  these  measures  (calculated  at  30  m)  (X30) 
to  develop  stage  I  models  for  predicting  the  slope  (pi)  and  intercept  (po)  of  the 
proportion  transition  lines  for  the  calibration  site;  c)  predict  30  m  proportions 
(Pq)  for  the  calibration  site  by  applying  the  stage  II  model  in  Eq.  2  using  the 
measured  1020  m  proportions  (Pr),  and  the  slopes  ($\)  and  intercepts  ($q) 
estimated  from  the  stage  I  models;  d)  using  the  stage  I  models  developed  on  the 
calibration  site,  repeat  step  c  and  evaluate  the  procedure  when  applied  to  the  data 
from  the  validation  site;  e)  invert  and  evaluate  the  procedure  by  running  the  stage 
I  models  using  the  spatial  variables  as  measured  at  1020  m  (Xio2o)>  and  supply 
the  results  to  the  stage  II  model  to  predict  30  m  proportions  based  only  on  infor- 
mation measured  at  1020  m  for  the  validation  site. 


METHODS  and  RESULTS 

The  Plumas  and  Stanislaus  National  Forests  are  used  as  calibration  and  vali- 
dation sites,  respectively.  Landsat  Thematic  Mapper  data  have  been  classified  to 
produce  maps  of  general  land-cover  categories  for  each  site.  Classes  include  bar- 
ren, brush,  hardwood,  water,  and  conifer.  Although  the  two  sites  have  similar 
characteristics,  they  are  spatially  separated  by  roughly  2°  of  latitude. 

The  data  for  each  site  are  aggregated  to  1020  m  resolution  using  a  plurality- 
based  aggregation  procedure.  This  involves  coding  each  grid  cell  in  a  1020  m 
sampling  grid  with  the  most  frequently  occurring  subgrid-cell  class. 

A  set  of  randomly  located  238x238  pixel  subregions  serve  as  the  sampling 
units  for  the  analyses.  Each  unit  contains  56,644  30  m  pixels  and  49  1020  m  pix- 
els. The  Plumas  contains  fifty  sampling  units  for  model  development  and  initial 
testing.  The  Stanislaus  contains  thirty-five  units  for  model  validation  and  model 
inversion.  The  number  of  units  represent  30%  of  all  possible  such  units  from 
each  site.  Within  each  sampling  unit  the  following  measurements  are  collected: 
proportions  at  30  m  for  each  class,  proportions  at  1020  m  for  each  class,  a  set  of 
spatial  measures  at  30  m,  and  the  same  set  of  spatial  measures  at  1020  m.  The 


85 


ultimate  goal  is  to  estimate  30  m  proportions  by  supplying  1020  m  area  measure- 
ments and  slope  and  intercept  coefficients  to  the  stage  II  model  (Eq.  2).  An  inter- 
mediate goal  is  to  estimate  the  proper  slopes  and  intercepts  using  a  multiple 
regression  model  with  a  parsimonious  set  of  spatial  measures  as  the  independent 
variables.  The  slopes  and  intercepts  of  the  proportion  transition  lines  are  the 
dependent  variables  in  these  stage  I  models. 

A  variety  of  spatial  measures  are  determined  within  each  unit  using  the  r.le 
software  (Baker  and  Cai  1992).  Five  of  these  demonstrate  scale-invariance  as 
determined  by  examining  the  simple  correlations  between  each  variable  and  itself 
at  the  two  different  scales.  Of  these  five,  three  prove  significant  in  modeling  both 
the  slope  and  the  intercept  of  the  proportion  transitions  as  determined  within  each 
sampling  unit.  An  additional  variable  (c  below)  that  does  not  have  the  scale 
invariance  property  is  included  because  it  characterizes  an  important  landscape 
characteristic  not  previously  included  in  the  model.  The  four  variables  used  are: 
maximum  class  size  (mx),  inverse  Simpson's  index  (s*1),  contagion  (c),  and 
entropy  (ent).  Expressions  for  the  latter  3  variables  are: 


Pi2 


c  =2-]n(k)-ent  (4) 


ent  =  - 


where  k  is  the  number  of  classes  present,  Pi  is  the  proportion  of  class  i  in  the 
sampling  unit,  and  Pn  are  elements  of  a  k  xk  co-occurrence  matrix  and  represent 
adjacency  probabilities  between  classes  i  and  j .  Maximum  class  size  refers  to 
the  proportion  of  the  largest  class  in  the  sampling  unit.  Simpson's  index  indicates 
the  probability  of  randomly  selecting  two  pixels  of  the  same  attribute.  Contagion 
measures  the  degree  of  clumping  in  the  landscape.  Entropy  is  maximized  when 
all  pixels  of  a  given  class  are  as  far  away  from  one  another  as  possible.  Table  1 
shows  the  cross-scale  correlation  matrix  for  this  set  of  variables. 

Table  2  presents  regression  summaries  for  the  two  stage  I  models.  The 
independent  variables  of  are  the  spatial  variables  measured  at  30  m  as  described 
above.  The  dependent  variables  are  a)  the  intercepts  and  b)  the  slopes  of  the  pro- 
portion transition  lines  as  determined  using  a  linear  least  squares  fit  between  the 
30  and  1020  m  proportions  for  the  classes  existent  within  each  sampling  unit. 


Table  1. — Cross-scale  correlations  (scale-invariance)  of  independent  variables. 


mxioio 

S  ll020 

c  1020     ent  1020 

mx3o 

0.91 

0.69 

C30 

-0.07 

ent30 

0.81 

86 


Table  2.— Stage  I  models.  Slope  model  R2adj =0.63.  Intercept  model  R2adj=0J0. 


Slope  Model 

Coefficient 

Standard  Error 

t-value 

P>|t| 

-5.77 

1.05 

-5.49 

0.00 

4.95 

0.84 

5.92 

0.00 

S~l30 

-0.74 

0.22 

-3.33 

0.002 

-0.679 

0.17 

-4.06 

0.0002 

ent30 

2.83 

0.49 

5.83 

0.00 

Intercept  Model 

Po 

0.0042 

0.093 

0.045 

0.96 

01X30 

-0.30 

0.074 

-4.054 

0.0002 

■S_130 

0.067 

0.020 

3.3904 

0.0015 

^30 

0.113 

0.015 

7.68 

0.00 

entw 

-0.145 

0.043 

-3.39 

0.0015 

The  stage  I  models  described  in  Table  2  estimate  the  $0  and  $\  coefficients 
necessary  for  employing  the  stage  II  model  (Eq.  2).  This  sequential  modeling 
process  is  conducted  three  times.  First,  the  stage  I  models  are  developed  using 
the  data  from  the  Plumas  (calibration  site).  The  predicted  values  from  these 
models  are  then  used  in  Eq.  2  to  test  the  overall  modeling  process  for  the  calibra- 
tion data.  Second,  the  stage  I  models  are  applied  in  a  predictive  mode  using  the 
independent  variables  as^  measured  at  30  m  from  the  Stanislaus  (validation  site). 
Again  predicted  po  and  Pi  values  supply  the  coefficients  to  run  Eq.  2  and  estimate 
30  m  proportions  for  the  Stanislaus.  Third,  stage  I  models  are  applied  using  the 
variables  measured  at  1020  m  from  the  Stanislaus  and  the  results  are  again  used 
to  estimate  30  m  proportions.  The  first  two  cases  are  forward  models  in  the  sense 
that  they  require  high  resolution  information  to  perform  the  correction.  The  third 
is  an  inverted  model,  because  it  relies  only  coarse  resolution  data.  The  results 
from  these  three  tests  are  presented  in  Figures  2,  3,  and  4. 

DISCUSSION  and  CONCLUSIONS 

As  seen  in  Figure  1,  Figures  2a,  3a,  and  4a  also  illustrate  the  basic  scaling 
effects  for  class  proportions.  Note  that  at  1020  m,  the  greatest  underestimations 
occur  for  intermediate-small  classes,  and  the  greatest  overestimations  occur  for 
intermediate-large  classes.  Very  large,  very  small,  and  moderate  sized  classes  (at 
the  cross-over  point,  around  30%)  are  all  reasonably  estimated  at  1020  m. 

The  goal  of  the  two  stage  modeling  procedure  is  to  pull  the  coarse-resolution 
area  estimates  closer  to  the  zero-one  line.  Figures  2b  and  3b  demonstrate  that  the 
model  improves  area  estimates  for  both  the  calibration  and  the  validation  sites 
when  operated  in  the  forward  mode.  The  correction  procedure  performs  best  for 
large  classes.  For  small  classes  a  notable  dip  (albeit  reduced)  below  the  zero-one 
line  still  occurs.  An  intercept  effect  is  also  evidenced  by  the  vertical  alignment  of 
estimates  above  the  zero  value  of  the  x-axis.  Results  from  the  inverted  model 
(Figure  4b)  also  show  general  improvement,  although  considerable  scatter  occurs 
for  very  large  classes  (note  the  2  outliers). 

Once  corrected  values  are  derived,  it  is  possible  to  tabulate  the  total  absolute 
error  within  each  sampling  unit.  For  a  given  sampling  unit  g ,  the  total  error  is: 


87 


Plumas 


Forward  Model:  Plumas 


0.2  0.4  0.6 

Measured  Proportions  at  30  m 


0.8 


a 

o 
E 


0.2  0.4  0.6 

Measured  Proportions  at  30m 


0.8 


Forward  Model:  Total  Error  by  Region 


0.3        0.4  0.5 
Pre-Correction  Error 


0.6 


0.7 


(d) 


Ho:  u1=u2 
T=4.47 
P>T:  0.00 


Pre-Correction 


Post-Correction 


Figure  2. — Forward  model  results  for  the  calibration  site. 


Stanislaus 


Forward  Model:  Stanislaus 


0.0  0.2  0.4  0.6  0.8 

Measured  Proportions  at  30  m 


o 
c 

a 

2 

0- 

E 

o 
o 


0.0  0.2  0.4  0.6  0.8 

Measured  Proportions  at  30m 


Forward  Model:  Total  Error  by  Region 


Pre-Correction  Error 


uj  n 


(d) 


Ho:u1=u2 
T=4.87 
P>T:  0.00 


Pre-Correction 


Post-Correction 


Figure  3. — Forward  model  results  for  the  validation  site. 


88 


error g  =  £  I P/.  1020  "  ^.30 1 

where  /V1020  ^'.30  are  tne  proportions  for  class  i  at  1020  and  30  m,  respec- 
tively. By  calculating  these  values  for  both  pre-  and  postcorrection  data,  the 
results  can  readily  be  compared.  Figures  2c  and  3c  show  the  relationship  between 
pre-  and  postcorrection  error  for  the  calibration  and  validation  sites  using  the  for- 
ward model.  For  any  point  falling  below  the  zero-one  line,  the  total  error  is 
reduced  due  to  the  correction  procedure.  At  both  sites,  the  total  error  for  roughly 
90%  of  the  sampling  units  is  either  reduced  or  unchanged  after  correction.  Con- 
versely, the  error  for  roughly  10%  of  the  units  is  increased.  For  the  inverted 
model  (Figure  4c)  80%  of  the  regions  are  either  improved  or  unchanged  after 
correction  with  two  positive  outliers  (regions  21  and  25). 

Figures  2d,  3d,  and  4d  are  boxplots  comparing  the  distributions  of  the  pre- 
and  postcorrection  total  error  values.  In  all  cases,  the  correction  procedure  results 
in  a  significant  reduction  in  error.  For  the  inverted  model  the  test  was  performed 
after  removing  the  outliers.  In  each  case,  however,  the  T-tests  are  suspect  due  to 
unequal  variance. 

Several  interesting  questions  deserve  continued  attention.  How  do  the  spa- 
tial measures  used  govern  the  proportion  transitions?  Will  the  scale-invariance 
property  transfer  to  other  landscape  types? 


Stanislaus 


Inverted  Model:  Stanislaus 


0.2  0.4  06  08 

Measured  Proportions  at  30  m 


a 

o 
0. 

E 

8 


0.0  02  0.4  06  0.8 

Measured  Proportions  at  30m 


Inverted  Model:  Total  Error  by  Region 


0  2 


03  0.4 
Pre-Correction  Error 


0.5 


Inverted  Model:  Two  Outliers  Removed 


(d) 


Ho:  u1=u2 
T=2.27 
P>T  0  027 


Pre-Correction 


Post-Correction 


Figure  4. — Inverted  model  results  for  the  validation  site. 


89 


Might  other  scale-invariant  measures  better  predict  proportion  transitions?  Is  the 
general  procedure  extensible  across  landscape  types?  What  is  the  effect  of  con- 
straining the  intercept  to  zero?  What  are  the  sensitivities  of  the  cross -over  point 
seen  in  Figure  1?  Resolving  these  and  other  issues  will  help  formalize  a  body  of 
understanding  of  how  class  proportions  scale.  This  understanding  hopefully  will 
lead  to  improved  land-cover  area  estimates  at  local,  regional  and  global  scales. 

REFERENCES 

Adams,  J.  B.,  Smith,  M.  O.,  and  Johnson,  P.  E.  1986.  Spectral  mixture  modeling: 
A  new  analysis  of  rock  and  soil  types  at  the  Viking  Lander  1  site.  /.  Geo- 
phy.Res.  91(B8):8098-8112. 

Baker,  W.  L.  and  Cai,  Y.  1992.  The  r.le  programs  for  multiscale  analysis  of 
landscape  structure  using  the  GRASS  geographical  information  system. 
Landscape  Ecology  7 (4): 29 1-302. 

Brown,  P.  J.  1982.  Multivariate  calibration.  /.  Royal  Statistical  Soc.  3:287-321. 

Cullinan,  V.  I.  and  Thomas,  J.  M.  1992.  A  comparison  of  quantitative  methods  for 
examining  landscape  pattern  and  scale.  Landscape  Ecology  7(3):21 1-227. 

Czaplewski,  R.  L.  and  Catts,  G.  P.  1992.  Calibration  of  remotely  sensed  propor- 
tion or  area  estimates  for  misclassification  error.  Remote  Sens.  Environ. 
39:29-43. 

Kalkhan,  M.  A.,  Reich,  R.  M.,  and  Czaplewski,  R.  L.  1995.  Evaluation  of  statisti- 
cal properties  of  the  inverse  estimator  for  remotely  sensed  areal  estimates 
using  simple  random  sampling.  Proc.  Amer.  Soc.  Photogramm.  and  Remote 
Sensing  Conf.,  27  Feb.  -  2  Mar.  1995,  Charlotte,  NC,  3:258-270. 

Legendre,  P.  and  Fortin,  M-J.  1989.  Spatial  pattern  and  ecological  analysis. 
Vegetatio  80:107-138. 

Mayaux  P.,  and  Lambin,  E.  F.  1995.  Estimation  of  tropical  forest  area  from 
coarse  spatial  resolution  data:  A  two  step  correction  function  for  propor- 
tional errors.  Remote  Sens.  Environ.  53:1-16. 

Moody,  A.  and  Woodcock,  C.  E.  1995.  The  influence  of  scale  and  the  spatial 
characteristics  of  landscapes  on  land-cover  mapping  using  remote  sensing. 
Landscape  Ecology  10(6):363-379. 

Moody,  A.  and  Woodcock,  C.  E.  1994.  Scale-dependent  errors  in  the  estimation 
of  land-cover  proportions  — Implications  for  global  land-cover  datasets. 
Photogramm.  Eng.  Remote  Sens.  60(5):585-594. 

BIOGRAPHICAL  SKETCH 

Aaron  Moody  is  a  geographer  at  the  University  of  North  Carolina,  Chapel 
Hill  with  a  specialization  in  remote  sensing  of  vegetation.  He  holds  a  Ph.D.  from 
Boston  University  and  an  M.A.  from  the  University  of  California  at  Santa  Bar- 
bara, both  in  geography. 


90 


Geostatistical  Analysis  of 
Multi-Spatial  Resolution  Imaging 

Spectrometer  Data  for 
Characterizing  Forest  Ecosystems 

Paul  M.  Treitz1  and  Philip  J.  Howarth2 

Abstract. — Multi-spatial  resolution  Compact  Airborne  Spectrographic 
Imager  (CASI)  reflectance  data  are  analyzed  using  geostatistical 
techniques.  Semivariograms  are  derived  from  visible  and  near- 
infrared  data  collected  at  two  altitudes  for  selected  forest  ecosystems 
representative  of  the  boreal  landscape  for  northwestern  Ontario, 
Canada.  These  analyses  reveal  that  (i)  mean  ranges  differ  for 
ecosystems  of  different  complexities;  (ii)  mean  ranges  for  the  near- 
infrared  data  are  greater  than  the  corresponding  ranges  derived  for  the 
visible  data;  (iii)  mean  ranges  are  greater  for  the  medium-altitude  data 
than  for  the  low-altitude  data;  (iv)  the  height  of  the  semivariogram 
(sill)  increases  for  stands  demonstrating  complex  layering,  particularly 
for  the  near-infrared  data;  and  (v)  the  sills  are  typically  lower  for  the 
medium-altitude  data.  These  results  indicate  that  for  optimal 
discrimination  of  forest  ecosystems,  multi-spatial  resolution  remote 
sensing  data  and  spatial  processing  techniques  offer  potential  for 
classification  and  mapping  of  forest  ecosystems.  


INTRODUCTION 

The  development  of  the  Forest  Ecosystem  Classification  (FEC)  for  northwestern 
Ontario,  Canada  was  spawned  by  the  need  to  produce  an  ecologically  based  forest 
classification  which  incorporates  both  physiographic  and  biotic  elements  of  the 
forest  ecosystem.  These  detailed  hierarchical  classifications  are  designed  for  field- 
level  mapping  and  are  therefore  difficult  to  implement  for  large  tracts  of  forest  that 
are  characteristic  of  much  of  northwestern  Ontario.  Remote  sensing  data  and  digital 
image  analysis  techniques  offer  potential  for  assisting  in  the  analysis  of  large  tracts 
of  forest  for  identification  of  relevant  ecosystem  classes,  particularly  in  the  context 
of  an  hierarchical  classification  scheme. 

However,  it  is  difficult  to  identify  appropriate  spatial  resolutions  for  remote  sensing 
data  acquisition  that  will  provide  the  most  suitable  level  of  information  for 
extracting  forest  ecosystem  parameters  for  a  variety  of  environments.  In  this  paper, 
the  results  of  geostatistical  analyses  of  multi-resolution  Compact  Airborne 
Spectrographic  Imager  (CASI)  reflectance  data  for  selected  forest  ecosystems  in 
northwestern  Ontario,  Canada  are  presented.  Using  these  types  of  analyses  for 
various  forest  ecosystems  at  different  spatial  resolutions,  it  is  possible  to  gain 
insight  into  the  impact  of  spatial  structure  on  information  extraction  from  these  data. 


Lecturer.  Department  of  Geography.  York  University:  North  York.  Ontario.  Canada 
Professor,  Department  of  Geography.  University  of  Waterloo.  Waterloo.  Ontario.  Canada 


91 


GEOSTATISTICAL  ANALYSIS  OF 
REMOTE  SENSING  DATA 

Jupp  etal.  (1988)  describe  many  remote  sensing  scenes  as  spatial  arrangements  of 
two-  or  three-  dimensional  objects  superimposed  on  a  uniform  background.  In  this 
scenario,  they  consider  a  discrete  object  scene  model  as  an  appropriate  abstraction 
of  the  scene.  In  this  model,  one  or  more  classes  of  objects  can  be  described  by  a 
unique  set  of  properties  or  parameters.  This  abstraction  is  appropriate  for  certain 
applications  of  remote  sensing  and  digital  image  analysis.  However,  within  a  forest 
ecosystem  context,  not  only  does  the  canopy  structure  vary  with  ecosystem  type, 
but  the  background  can  also  vary  significantly  with  respect  to  composition  and 
structure. 

The  brightness  values  of  a  remotely  sensed  image  are  a  function  of  spatial  position 
during  image  acquisition.  In  the  realm  of  geostatistics,  these  brightness  values  can 
therefore  be  considered  as  values  of  a  "regionalized  variable"  (Matheron,  1963).  To 
apply  regionalized  variable  theory  to  the  analysis  of  remotely  sensed  data,  adoption 
of  a  stochastic  view  of  the  landscape  and  its  spatial  structure  is  required  (Jupp  et  ai, 
1988).  This  is  a  logical  assumption  since  underlying  processes  and  properties  of  the 
landscape  will  produce  many  similar  scenes.  In  fact,  this  is  intrinsic  to  the 
classification  of  ecosystems.  It  is  assumed  that  similar  ecosystems  will  arise  from 
similar  environmental  conditions  and  processes  linked  closely  to  landscape.  In  this 
study,  semivariogram  analysis  is  used  to  examine  the  spatial  structure  of  high 
resolution  CASI  data  collected  for  natural  forest  ecosystems,  characteristic  of 
northwestern  Ontario. 

Semivariogram  Analysis 

The  semivariogram,  which  originates  from  the  theory  of  regionalized  variables 
developed  by  Matheron  (1963),  is  used  to  measure  the  spatial  dependence  of 
neighboring  observations  for  any  continuously  varying  phenomenon.  Hence,  it  is  a 
technique  that  can  be  applied  to  spectral  data,  a  variable  for  which  position  in  space 
and  time  is  known.  In  this  manner,  spatial  variation  in  images  can  be  examined  in 
relation  to  ground  scene  and  sensor  parameters  (Woodcock  et  ai,  1988). 

The  semivariogram  plots  semivariance  against  spatial  separation  along  a  given 
relative  orientation  and  provides  a  concise  and  unbiased  depiction  of  the  scale  and 
pattern  of  spatial  variability  (Curran,  1988).  The  semivariogram  or  y(h)  is 
calculated  as: 

y(h)  =  — !— -  X[Z(*« )  -     +  W2 

2{n-h)  I=1 

where  h  is  the  lag  (or  distance)  over  which  y  (semivariance)  is  measured,  n  is  the 
number  of  observations  used  in  the  estimate  of  y(h),  and  Z  is  the  value  of  the 
variable  of  interest  at  spatial  position  xt  .  The  value  Z(*;  +  h)  is  the  variable  value  at 
distance  h  from  x,\  In  this  study,  y(h)  estimates  the  variability  of  reflectance,  Z,  as  a 
function  of  spatial  separation. 


92 


The  semivariogram  has  proven  useful  in  remote  sensing  because  it  enables 
researchers  to  relate  some  of  the  descriptors  of  the  semivariogram  to  the  spatial 
characteristics  of  the  scene.  For  example,  the  range,  which  defines  the  distance  at 
which  pixels  are  not  spatially  related,  provides  a  measure  of  the  size  of  the  elements 
in  the  scene  and  has  been  suggested  as  a  useful  indicator  in  selecting  the  optimal 
spatial  resolution  for  discriminating  the  features  embedded  in  the  image 
semivariogram  (Quran,  1988).  Woodcock  et  al.  (1988)  also  found  that  the  density 
of  coverage  of  objects  in  the  scene  affects  the  height  of  the  semivariogram. 
Atkinson  and  Danson  (1988)  used  semivariograms  to  measure  spatial  dependence 
in  coniferous  and  oak  plantations.  They  found  the  range  of  the  semivariogram  was 
related  to  stand  age  and  species,  and  were  able  to  determine  the  optimal  spatial 
resolutions  for  even-aged  stands.  Cohen  et  al.  (1990)  found  the  ranges  for  lm 
spatial  resolution  data  were  related  to  the  mean  tree  canopy  sizes  of  the  stands. 

STUDY  AREA 

The  study  area  is  located  within  the  Central  Plateau  section  of  the  Boreal  Forest 
Region  (Rowe,  1972)  approximately  100  km  north  of  Thunder  Bay,  Ontario, 
Canada.  A  diverse  mosaic  of  forest  stand  types  with  various  soil  and  landform 
conditions  resulting  from  glacial  activities  are  characteristic  of  the  area.  This  area 
represents  a  low-  to  moderate-relief  boreal  environment  with  a  rolling  topography 
and  terrain  controlled  by  bedrock.  Trembling  aspen  {Populus  tremuloides)  and 
black  spruce  {Picea  mariana)  are  dominant  with  jack  pine  {Pinus  banksiana),  white 
spruce  {Picea  glauca),  balsam  fir  {Abies  balsamea),  white  birch  {Betula  papyrifera), 
white  cedar  {Thuja  occidentalis)  and  tamarack  {Larix  laricina)  occurring  in  various 
mixtures.  Forest-stand  overstories  are  monospecific  or  mixed  and  understories 
range  from  shrub-  and/or  herb-rich  to  poor. 

The  FEC  that  is  applied  in  the  study  area  is  the  Northwestern  Ontario  (NWO)  FEC 
(Sims  et  al,  1989).  The  framework  upon  which  FECs  are  based  incorporates  those 
components  of  forest  site  which  contribute  to  forest  development  (i.e.,  canopy  and 
understory  vegetation,  soils,  landform,  climatic  regime,  and  regional  physiography) 
(Sims  and  Uhlig,  1992).  FECs  were  developed  for  stand-level  (i.e.,  <1: 10,000) 
application  to  provide  information  regarding  vegetation,  soil  and  site  conditions. 
The  basic  units  of  FECs  are  'Vegetation  Types'  (V-Types)  and  'Soil  Types'  (Sims 
et  al,  1989).  Since  FECs  are  hierarchical,  field-level  units  can  be  aggregated  to 
create  lower-resolution  ecosystem  units  (Hills  and  Pierpoint,  1960). 

METHODS 

Remote  Sensing  Data  Acquisition 

CASI  data  were  acquired  over  the  study  site  on  July  30,  1993.  These  data  were 
collected  at  approximately  10:30  a.m.  along  a  flight  line  oriented  parallel  to  the  solar 
azimuth  in  order  to  minimize  the  effects  of  bidirectional  reflectance  (BDRF).  A 
nine-band  multispectral  dataset  was  collected  at  two  different  altitudes  with  spatial 
resolutions  of  1.39m  x  5.36m  (figures  1  and  2);  and  0.73m  x  5.36m  (figures  3  and 
4).  The  CASI  data  were  converted  to  radiance  and  then  to  reflectance  using 
procedures  outlined  in  Babey  and  Soffer  (1992)  and  Shepherd  (1994). 


93 


Semivariance  Analysis  of  Multi-Spatial  Resolution 
Remote  Sensing  Data 

In  this  study,  experimental  semivariograms  are  used  to  model  the  underlying 
semivariogram  for  the  CASI  reflectance  data.  A  single  visible  band  and  a  near- 
infrared  band  were  used  for  calculating  semivariance  and  deriving  semivariograms 
from  low-  and  medium-altitude  CASI  data.  Homogeneous  stands  (landscape  units) 
of  sufficient  size  were  identified  on  aerial  photographs  and  located  on  the  low-  and 
medium-altitude  CASI  images.  To  examine  the  spatial  variability  of  forest 
ecosystems  at  high  spatial  resolutions,  transects  of  100  pixels  in  the  cross-track 
direction  were  extracted  from  the  CASI  data  for  selected  forest  stands  representing 
individual  V-Types  or  complexes  of  V-Types  (table  1).  These  stands  are  arranged 
in  the  table  and  in  the  corresponding  figures  along  a  gradient  of  pure  hardwood  to 
pure  conifer.  These  visible  and  near-infrared  data  were  subjected  to  semivariogram 
analysis1  to  estimate  range  and  sill  values  using  a  spherical  model.  A  series  of 
transects  were  analyzed  for  each  forest  stand,  from  which  the  mean  ranges  and  sills 
were  calculated  to  characterize  the  stand  /  landscape  unit. 


Figure  1.— Medium-altitude  data  Figure  2.— Medium-altitude  data 


Figure  3.— Low-altitude  data  Figure  4.— Low-altitude  data 


1  Geostatistical  software:  VARIOWIN  (V.2.1)  developed  by  Y.  Pannatier  at  the  Institute  of  Mineralogy  of  the  University 
of  Lausanne  SWITZERLAND  (PANNATIER,  Y.,1994.  MS-WINDOWS  for  Exploratory  Variography  and  Variogram 
Modeling  in  2D.  In  Proceedings,  Statistics  of  Spatial  Processes:  Theory  and  Applications.  Capasso,  V.,  Girone  G.  and 
Posa,  D.  Eds.,  Bari,  Italy,  Sep.  27-30,  1993  pp.  165-170). 


94 


Table  1.  —  Forest  ecosystem  groupings  for  semivariogram  analysis 


Forest  Ecosystem  V-Types  (defined  by  Sims  et  a/.,  1989) 
Grouping  


Aspen 

V5 

Aspen  Hardwood 

vo 

Trembling  Aspen  (W.  Birch)  -  Balsam  Fir  /  Mountain  Maple 

Vo 

Trembling  Aspen  (W.  Birch)  /  Mountain  Maple 

Aspen  Complex 

V9 

Balsam  Fir  -  White  Spruce  Cornier  and  Mixedwood 

V10 

Trembling  Aspen  -  Black  Spruce  -  Jack  Pine  /  Low  Shrub 

VI 1 

Trembling  Aspen  -  Conifer  /  Blueberry  /  Feathermoss 

Aspen  Mixedwood 

V8 

Trembling  Aspen  (White  Birch)  /  Mountain  Maple 

VI 1 

Trembling  Aspen  -  Cornier  /  Blueberry  /  Feathermoss 

V1Q 

V  17 

DititK  opruLe  ivuxcuwoou  /  neru  ivicn 

Conifer  Mixedwood 

Vo 

Trembling  Aspen  (White  Birch)  /  Mountain  Maple 

\/ 1  A 

V  14 

Balsam  Fir  Mixedwood 

V15 

White  Spruce  Mixedwood 

V  1 0 

joaisam  rir  -  wnne  opruce  iviixeowooo  /  reatnermoss 

V 1  Q 
V  1  y 

dijlk  opruce  Mixeowooo  /  rierD  Kicn 

Eastern  White  Cedar 

V22 

Eastern  White  Cedar 

Upland  Black  Spruce 

V 55 

Black  Spruce  /  Feathermoss 

Black  Spruce  Mix 

V  55 

Black  Spruce  /  Feathermoss 

V34 

Black  Spruce  /  Labrador  Tea  /  Feathermoss  (Sphagnum) 

V35 

Black  Spruce  /  Speckled  Alder  /  Sphagnum 

V36 

Black  Spruce  /  Bunchberry  /  Sphagnum  (Feathermoss) 

Lowland  Black  Spruce 

V35 

Black  Spruce  /  Speckled  Alder  /  Sphagnum 

V36 

Black  Spruce  /  Bunchberry  /  Sphagnum  (Feathermoss) 

V37 

Black  Spruce  /  Ericaceous  Shrub  /  Sphagnum 

Wetland  Black  Spruce 

V37 

Black  Spruce  /  Ericaceous  Shrub  /  Sphagnum 

V38 

Black  Spruce  /  Leatherleaf  /  Sphagnum 

RESULTS 


At  the  low  altitude,  and  for  the  stands  sampled,  the  mean  ranges  derived  from  the 
semivariograms  for  the  visible  band  indicate  that  trembling  aspen  stands  have 
greater  ranges  (=  8  pixels  /  5.8m)  )  than  upland  black  spruce  stands  (=  6  pixels  / 
4.4m)  and  lowland  black  spruce  stands  (=  4  pixels  /  2.9m)  (figure  5).  In  general, 
the  range  values  for  the  visible  band  increase  with  stand  complexity.  A  similar  trend 
is  observed  with  respect  to  the  range  values  for  the  near-infrared  data  whereby 
hardwood  and  conifer  mixedwood  stands  have  greater  ranges  than  lowland  black 
spruce.  For  example,  range  values  are  approximately  10  pixels  /  7.3m  for  aspen 
complexes;  1 1  pixels  /  8.0m  for  conifer  mixedwood;  9  pixels  /  6.6m  for  upland 
black  spruce;  and  6  pixels  /  4.4m  for  lowland  black  spruce.  The  ranges  derived  for 
the  near-infrared  data  are  generally  greater  than  those  derived  for  the  visible  data  for 
each  ecosystem  class  (e.g.,  6  pixels  /  4.4m  (infrared)  versus  4  pixels  /  2.9m 
(visible)  for  lowland  black  spruce).  The  sills  for  the  trembling  aspen  and  conifer 
mixedwood  stands  are  much  greater  than  those  of  the  lowland  black  spruce  stands, 
particularly  for  the  near-infrared  data.  For  example,  the  mean  near-infrared 
semivariance  (sill)  is  approximately  1700  for  aspen  stands;  2200  for  conifer 
mixedwood  stands;  and  350  for  lowland  black  spruce  stands  (figure  6).  Although 
the  trend  is  similar  in  the  visible  band,  the  differences  are  not  as  extreme. 


95 


Forest  Ecosystem  Complexes 


Figure  5. — Mean  range  values  (580.06nm  -  600.94nm)  for  forest  ecosystem 
complexes  in  the  Rinker  Lake  study  area  (Altitude  =  600m) 


3000 


2500 


2000 


«  1500 
I 


1000 


500 


Forest  Ecosystem  Complexes 


Figure  6. — Semivariance  (743.64nm  -  750.26nm)  for  forest  ecosystem 
complexes  in  the  Rinker  Lake  study  area  (Altitude  =  600m) 


96 


For  the  medium-altitude  data,  the  differences  in  mean  ranges  between  the  aspen 
stands  and  the  lowland  black  spruce  stands  for  the  near-infrared  data  are  not  as 
pronounced.  However,  the  ranges  for  the  near-infrared  data  are  greater  than  those 
for  the  visible  data  for  all  stands,  particularly  lowland  black  spruce  (i.e.,  7  pixels  / 
9.7m  versus  3.5  pixels  /  4.9m).  Ranges  derived  from  the  medium-altitude  data  are 
greater  than  those  measured  for  the  low-altitude  data.  For  example,  the  mean  ranges 
for  the  trembling  aspen  stands  are  approximately  7.5  pixels  /  5.5m  for  the  low- 
altitude  data  while  for  the  medium-altitude  data  they  are  approximately  6.5  pixels  / 
9.0m.  Finally,  the  sills  are  typically  lower  for  the  medium-altitude  data  in 
comparison  to  the  low-altitude  data. 

CONCLUSIONS 

For  the  low-altitude  CASI  data,  the  hardwood  and  mixed  conifer  stands  have 
greater  range  values  than  pure  conifer,  particularly  lowland  conifer  where  canopy 
and  understory  are  generally  simple  and  homogeneous.  At  this  resolution, 
individual  trees  may  be  the  dominant  feature  affecting  the  semivariogram, 
particularly  for  the  black  spruce  and  trembling  aspen  stands  since  the  crown 
diameters  of  these  species  are  approximately  equal  to  the  ranges  of  influence. 
However,  since  these  pixels  are  not  square  and  individual  trees  are  not  observable, 
it  is  more  likely  that  the  ranges  of  influence  are  determined  by  an  integration  of  a 
uniform  number  of  tree  crowns  with  understory  and  ground  cover  components. 

The  contrasting  ranges  of  the  near-infrared  and  visible  data  indicate  that  different 
features/phenomena  are  being  measured  in  the  near  infrared  as  opposed  to  the 
visible,  or  at  least  the  proportional  contributions  of  features/phenomena  to 
reflectance  differ  between  these  data.  It  is  likely  that  the  near-infrared  data  contain 
more  ground-cover  information  since  near-infrared  energy  has  greater  potential  for 
penetration  through  a  forest  canopy.  The  differences  in  the  range  values  observed  at 
each  spatial  resolution  for  the  same  ecosystems  reveals  that  information  content  also 
differs  at  the  two  altitudes.  This  observation  would  indicate  that  different 
features/phenomena  or  aggregations  of  features/phenomena  (e.g.,  possibly  2-3  tree 
crowns  with  associated  understories)  are  being  measured  at  the  medium  resolution. 

The  observed  differences  in  the  magnitude  of  semivariance  at  both  resolutions 
indicates  that  the  trembling  aspen  stands  (pure  and  mixed)  and  conifer  mixedwood 
stands  have  increased  layering  with  a  greater  volume  of  total  cover  than  the  black 
spruce  stands.  The  complexity  of  the  stand  is  more  prevalent  with  the  near-infrared 
data,  since  near-infrared  energy  has  a  greater  capability  of  interacting  with 
understory  components.  The  occurrence  of  lower  sills  for  the  medium-altitude  data 
indicates  increased  regularization  as  a  result  of  decreased  variability  of  forest  stand 
reflectance  at  this  resolution. 

Based  on  these  observations,  it  is  anticipated  that  the  potential  for  identifying  forest 
ecosystems  at  a  variety  of  levels  will  be  improved  by  applying  spatial  analysis 
techniques  (e.g.,  texture  processing,  spatial  interpolation)  to  multi-spatial  resolution 
remote  sensing  data.  The  identification  of  a  single  optimal  spatial  resolution  for 
forest  ecosystem  discrimination  may  not  be  possible.  Instead,  it  may  be  more 
suitable  to  analyze  multi-spatial  resolution  data  with  appropriate  processing 
techniques  to  extract  various  scales  of  information. 


97 


ACKNOWLEDGMENTS 


Funding  for  this  project  has  been  provided  through  the  Northern  Ontario 
Development  Agreement,  Northern  Forestry  Program.  Additional  support  has  been 
provided  by  a  Center  of  Excellence  Grant  from  the  Province  of  Ontario  to  the 
Institute  for  Space  and  Terrestrial  Science. 

REFERENCES 

Atkinson,  P.  and  Danson,  F.  1988.  Spatial  resolution  for  remote  sensing  of  forest 

plantations,  IGARSS'88  Symposium,  Edinburgh,  Scotland,  ESA  Publications 

Division,  pp.  221-223. 
Babey,  S.  and  Soffer,  R.  1992.  Radiometric  calibration  of  the  Compact  Airborne 

Spectrographic  Imager  (CASI),   Canadian  Journal  of  Remote  Sensing, 

18(4):233-242. 

Curran,  P.J.  1988.  The  semivariogram  in  remote  sensing:  an  introduction,  Remote 

Sensing  of  Environment,  24:  493-507. 
Hills,  G.  and  Pierpoint,  G.  1960.  Forest  site  evaluation  in  Ontario,  Research 

Report  No.  42,  Ontario  Department  of  Lands  and  Forests,  Toronto,  Ontario, 

Canada,  63  pp. 

Jupp,  D.L.B.,  Strahler,  A.H.  and  Woodcock,  C.E.  1988.  Autocorrelation  and 

regularization  in  digital  images:  I.  basic  theory,  IEEE  Transactions  on 

Geoscience  and  Remote  Sensing,  26(4):463-473. 
Matheron,  G.  1963.  Principles  of  geostatistics,  Economic  Geology  58:1246-1266. 
Rowe,  J.  1972.  Forest  Regions  of  Canada,  Publication  No.  1300,  Canadian 

Forestry  Service,  Department  of  the  Environment,  Ottawa,  Ontario,  Canada, 

172  pp. 

Shepherd,  P.  1994.  Section  3.3,  Airborne  Remote  Sensing  Data,  in  Kalnins,  V., 
Treitz  P.  and  Howarth,  P.,  Rinker  Lake  Data  Report:  1993-1994,  Earth- 
Observations  Laboratory,  Technical  Report  ISTS-EOL-TR94-002,  pp.  19-48. 

Sims,  R.  and  Uhlig,  P.  1992.  The  current  status  of  forest  site  classification  in 
Ontario,  The  Forestry  Chronicle,  68(l):64-77. 

Sims,  R.,  Towill,  W.,  Baldwin,  K.  and  Wickware,  G.  1989.  Field  Guide  to  Forest 
Ecosystem  Classification  for  Northwestern  Ontario,  Ontario  Ministry  of  Natural 
Resources,  Toronto,  Ontario,  Canada,  191  pp. 

Woodcock,  C.E.,  Strahler,  A.H.  and  Jupp,  D.L.B.  1988.  The  use  of  variograms 
in  remote  sensing:  I.  scene  models  and  simulated  images,  Remote  Sensing  of 
Environment,  25:323-348. 

BIOGRAPHICAL  SKETCH 

Paul  Treitz  is  a  Lecturer  in  the  Department  of  Geography  (Faculty  of  Arts)  and 
Faculty  of  Environmental  Studies  at  York  University,  North  York,  Ontario, 
Canada.  He  is  also  a  Ph.D.  student  in  the  Department  of  Geography  at  the 
University  of  Waterloo,  Waterloo,  Ontario,  Canada. 

Philip  Howarth  is  a  Professor  of  Geography  at  the  University  of  Waterloo, 
Waterloo,  Ontario,  Canada.  He  is  also  a  Principal  Investigator  with  the  Earth- 
Observations  Laboratory  of  the  Institute  for  Space  and  Terrestrial  Science,  a  center 
of  excellence  established  by  the  Province  of  Ontario  for  remote  sensing  and  space 
related  research. 


98 


Application  of  Multiscale  Modeling 
for  Characterizing  Variability  in 
Hydrologic  Processes 

Praveen  Kumar1 


Abstract. -We  describe  a  multiscale  modeling  framework  applicable 
for  a  wide  range  of  hydrologic  processes.  The  foundation  for  this  work 
has  been  laid  by  Basseville  et  al.  [1992]  and  Chou  et  al.  [1994].  Their 
development  is  based  on  treating  the  scale  parameter  akin  to  time, 
such  that  description  at  a  particular  scale  captures  the  features  of 
the  process  up  to  that  scale  that  are  relevant  for  the  prediction  of 
finer  scale  features,  as  in  the  notion  of  state-space  approaches.  We 
show  the  applicability  of  this  model  for  modeling  and  estimation  of  a 
broad  range  of  multiscale  phenomena  in  hydrology.  The  methodology 
is  illustrated  through  an  example  using  a  soil-moisture  field  obtained 
using  remote  sensing  observations. 


INTRODUCTION 

In  varied  guises  throughout  hydrologic  science  we  encounter  questions  con- 
cerning the  quantitative  relationship  between  processes  occurring  at  disparate 
spatial  and  temporal  scales.  These  questions  are  mostly  triggered  by  our  need 
to  understand,  model  and  predict  hydrologic  processes  over  a  broad  range 
of  spatial  and  temporal  scales.  Questions  about  multiscale  manifestation  of 
processes  are  at  the  heart  of  contemporary  hydrologic  research  and  answers 
to  these  questions  will  significantly  advance  our  efforts  in  understanding  the 
dynamics  of  the  large  scale  hydrologic  cycle. 

Over  the  last  several  years  there  has  been  significant  effort  devoted  to  or- 
ganizing field  experiments  in  combination  with  remote  sensing  measurements, 
such  as  FIFE  {Sellers  et  al.,  1992)  and  SIR-C  [Stofan  et  al.,  1995],  to  improve 
our  understanding  of  surface  hydrologic  and  ecological  processes  and  their  im- 
pact on  the  atmosphere.  In  order  to  improve  our  understanding  of  a  process 
at  a  particular  scale  we  need  to  make  measurements  at  that  scale.  Often  it 
is  difficult  to  decide  the  scale  of  significance  and  the  strategy  adopted  is  to 
make  measurements  at  several  scales  with  fine  scale  measurements  of  limited 
coverage  embedded  within  coarse  scale  measurements  of  larger  coverage  (see 
Figure  1).  Further  analysis  is  performed  to  integrate  the  information  across 
scales,  to  understand  what  area-averaged  fluxes  represent  in  conjunction  with 

department  of  Civil  Engineering,  University  of  Illinois,  Urbana,  Illinois  61801 


99 


point  fluxes,  and  what  physical  controls  on  area-averaged  fluxes  need  to  be 
considered  by  a  prognostic  model  [Smith  et  a/.,  1992]. 


i  <       Coarse  Scale  Grid 


Fine  Scale  Grid 

Figure  1:  Schematic  showing  an  idealized  embedding  of  fine  scale  measurement 
grid  of  limited  coverage  within  a  coarse  scale  grid  of  larger  coverage  for  a  two 
level  scheme. 

Different  techniques  exist  for  analyzing  observations  obtained  at  different 
resolutions  and  are  well  explained  in  Daley  [1991].  These  range  from  sim- 
ple function  fitting  and  stochastic  interpolation  (univariate  and  multivariate) 
schemes  to  four  dimensional  data  assimilation  (FDDA).  As  indicated  by  Daley 
[1991],  the  essence  of  a  good  scheme  is  to  account  for  the  following  factors 
simultaneously: 

1.  Observations  have  errors  that  can  be  temporally  and  spatially  correlated 
with  each  other. 

2.  The  process  state  variables  are  related  to  each  other  by  the  governing 
laws  and  are  not  independent. 

3.  There  can  be  large  data  voids  where  no  direct  observations  exist. 

An  ideal  algorithm  should  filter  observational  noise,  variance  associated  with 
scales  too  small  to  be  properly  resolved  by  the  sampling  scheme,  and  fluctua- 
tions that  do  not  satisfy  the  governing  laws. 

Although  multivariate  statistical  analysis  or  FDDA  address  these  issues, 
they  do  not  include  the  crucial  element  of  scale  as  an  explicit  parameter  in  the 
formulation.  This  causes  severe  limitations  in  our  ability  to  interpret  the  be- 
havior of  the  phenomenon  at  different  scales  and  the  influence  of  the  variability 
of  one  process  at  a  small  scale  on  the  variability  of  a  related  process  at  a  larger 
scale,  except  as  a  gross  aggregate.  For  examples,  Several  studies  (Mahrt  [1987], 
Wetzel  and  Chang  [1988],  Avissar  and  Pielke  [1989])  have  indicated  that  grid 


100 


scale  flux  averages  and  sub-grid  flux  variance  are  both  critical  in  characterizing 
the  response  of  the  atmosphere  to  surfaces  fluxes.  In  addition  no  multiscale 
modeling  scheme  is  immediately  available  from  these  formulations.  In  order 
to  address  these  issues  and  more  effectively  utilize  the  observations  obtained 
from  the  hierarchical  sampling  scheme  (see  Figure  1)  to  formulate  conjectures 
and  develop  models  we  need  an  analysis-modeling-assimilation  technique  that 
is  inherently  hierarchical  in  scale. 

Wavelet  transforms  have  been  significantly  exploited  in  geophysics  for  the 
analysis  of  processes  to  infer  their  multiscale  variability.  They,  however,  are 
not  easily  amenable  as  a  modeling  tool.  Motivated  by  the  scale  to  scale  hierar- 
chical decomposition  of  wavelet  transforms,  significant  advances  in  multiscale 
modeling  have  been  made  by  Basseville  et  al.  [1992]  and  Chou  et  al.  [1994]. 
The  key  to  their  development  is  the  treatment  of  the  scale  parameter  akin  to 
time,  such  that  description  at  a  particular  scale  captures  the  features  of  the 
process  up  to  that  scale  that  are  relevant  for  the  prediction  of  finer  scale  fea- 
tures, as  in  the  notion  of  state-space  approaches.  These  developments  provide 
a  natural  setting  not  only  for  analysis  and  modeling  of  multiscale  phenomena 
but  also  for  multiscale  data  assimilation. 

This  paper  attempts  to  provide  more  visibility  to  this  multiscale  modeling 
tool  in  the  geophysics  community  through  illustration  using  a  simple  example. 
In  the  next  section,  we  briefly  review  the  modeling  framework  and  then  provide 
an  application  to  modeling  soil  moisture  fields  and  estimating  missing  values. 

MULTISCALE  STATE-SPACE  MODEL 

The  following  description  briefly  summarizes  the  basics  of  the  ideas  developed 
by  Basseville  et  al.  [1992]  and  Chou  et  al.  [1992].  Consider  the  problem  of 
disaggregating  a  random  field  (in  two  dimensions)  from  coarse  to  fine  resolu- 
tion. At  the  coarsest  resolution  the  field  will  be  represented  by  a  single  value 
(see  Figure  2).  At  the  next  resolution  there  will  be  four  values  and  in  general, 
at  the  mth  resolution  we  obtain  4m  values.  The  values  of  the  random  field  can 
be  described  on  the  index  set  (m,i,j)  where  m  represents  the  resolution  and 
the  location  index.  The  scale  to  scale  decomposition  can  be  schematically 
depicted  as  a  tree  structure  (quadtree  for  two-dimensional  processes)  as  shown 
in  Figure  3.  To  describe  the  model  let  us  use  an  abstract  index  A  to  specify 
the  nodes  on  the  tree  and  let  7A  specify  the  parent  node  of  A  (see  Figure  3). 
Then  the  multiscale  stochastic  process  can  be  represented  as 

X(X)  =  A(X)X(jX)  +  B{\)W(\).  (1) 


101 


Resolution  1 


Resolution  2 


Figure  2:  The  structure  of  a  multiscale  random  field  is  shown.  The  values 
at  various  grid  locations  are  given  as  xm(i,j)  where  m  is  the  resolution 

index.  At  the  coarsest  resolution  (m  =  0),  the  field  is  represented  by  a  single 
value  or  state  vector,  and  generally  at  the  mth  resolution  there  are  4m  state 
vectors  (scale  and  resolution  are  inversely  related). 

A(X)  and  B(X)  are  matrices  of  appropriate  sizes.  We  can  assume  that  Xq  = 
X(0)  and  W  are  normally  distributed  with  known  covariance  matrices,  i.e., 

X0   ~   N(0,Po)      Po  =  E[X{0)XT(0)}  (2) 
W(X)   ~   N(0,I)  (3) 

where  /  is  the  identity  matrix.  Interpreting  the  states  at  a  given  level  of  the 
tree  as  a  representation  of  the  process  at  one  scale,  we  see  that  equation  (1) 
describes  the  evolution  of  the  process  from  coarse  to  fine  scale.  The  term 
A(\)X('j\)  represents  the  interpolation  or  prediction  down  to  the  next  finer 
level  and  B(X)W(X)  represents  new  information  added  as  the  process  evolves 
from  one  scale  to  the  next.  The  covariance  =  E[X(\)XT '(A)]  of  the  state 
at  node  A  evolves  as 

Px=A(\)PiXAT(\)  +  B(\)BT(\).  (4) 

Notice  that  the  form  of  equations  (1)  and  (4)  do  not  change  whether  we  deal 
with  one-dimensional  or  multi-dimensional  processes. 


102 


Figure  3:  Multiscale  decomposition.  The  abstract  index  A  refers  to  a  node  in 
the  tree  and  7A  refers  to  the  parent  node. 

A  particular  form  of  model  (1)  is  already  popular  as  disaggregation  model 
in  hydrologic  time  series  analysis  [see  Valencia  and  Schaake,  1973].  Equation 
(1),  however,  provides  a  more  general  modeling  framework.  For  instance,  it  is 
possible  to  choose  A(X)  and  B(X)  such  that  (1)  leads  to  a  fractal  model.  For 
example,  the  model 

X(X)  =  X(7A)  +  B4-p[o^  X/2W(X)       where  B  is  a  constant  (5) 

leads  to  fractal  processes  with  spectrum  1//^  [Chou  et  a/.,  1992]  ((3  =  2 
corresponds  to  Brownian  motion). 

Multiscale  stochastic  models  can  be  used  to  make  estimates  of  processes 
from  noisy  measurements  and  for  assimilation  of  data  obtained  at  different 
resolutions.  If  we  are  given  noisy  measurements  Y(X)  of  the  process  X(X)  we 
can  develop  an  estimation  problem  using  the  combination  of  equation  (1)  and 
the  following  equation 

Y{\)=.C(\)X(\)  +  V{\)  (6) 

where  V(X)  ~  N(0,  R(X)).  The  Matrix  C(A)  can  specify,  in  a  very  general 
way,  measurements  taken  at  different  spatial  locations  and  at  different  scales. 
R(X)  specifies  the  covariance  of  the  measurement  errors  V(A).  Equations  (1) 
and  (6)  can  be  solved  jointly  to  obtain  estimates  of  A'(A).  Notice  that  this  is 
a  very  attractive  technique  as  it  enables  us  to  combine  estimation  and  filtering 
while  exploiting  hierarchical  measurements  at  different  scales.  Also,  note  that 
coarse  scale  observations  provide  the  "context"  or  the  background  field  for  the 
next  finer  scale,  i.e..  scale  plays  a  role  akin  to  time  in  existing  data  assimilation 
schemes. 


103 


APPLICATION  TO  A  SOIL  MOISTURE  RANDOM  FIELD 


We  describe  the  applicability  of  the  multiscale  model  for  modeling  a  soil  mois- 
ture field  and  estimating  missing  values  using  a  fractal  prior  model.  The  data, 
collected  during  the  Washita'  92  Experiment  [Jackson  et  al.  1993],  were  ob- 
tained using  ESTAR  -  Electronically  Steered  Thinned  Array  Radiometer  - 
images  and  converted  to  volumetric  soil  moisture  content.  These  data  sets, 
which  correspond  to  several  days  of  daily  measurements  of  soil  moisture  in  the 
top  5  cm  layer,  show  scaling  or  fractal  characteristics  [Rodriguez- Iturbe  et  a/., 
1995].  We  study  the  field  obtained  on  the  second  day  of  measurement  (June 
11,  1992)  by  extracting  a  sub-field  with  grid  dimension  64  x  64  (Figure  4),  from 
the  original  of  228  x  93.  As  is  evident  from  the  Figure  there  are  several  missing 
values,  indicated  by  dark  patches,  either  due  to  the  presence  of  township  or 
roads,  or  measurement  errors.  For  this  study  we  do  not  distinguish  between 
these  and  treat  all  as  missing  values  due  to  measurement  errors. 

We  use  the  results  of  analysis  presented  by  [Rodriguez-Iturbe  et  a/.,  1995] 
to  estimate  the  parameter  (3  =  0.79  in  equation  5.  The  multiscale  model  we 
use  is  give  by  the  equations: 


where  V(X  =  6)  ~  N(0,cr2  =  25)  gives  the  noise  distribution  at  the  finest 
scale.  The  algorithm  first  estimates  the  coarser  scale  values  from  the  given 
finer  scale  values,  i.e.,  going  upward  in  the  tree  in  Figure  3,  and  after  reaching 
the  root  node  proceeds  downward  to  predict  the  estimates  at  all  grid  points 
and  all  scales,  thereby  providing  estimates  of  the  missing  values.  The  results 
of  the  multiscale  estimation  are  given  in  Figure  4.  As  is  evident,  the  algorithm 
does  an  extremely  good  job  of  estimating  the  missing  values.  The  estimation 
error  are  within  the  range  of  the  data  values. 

The  above  is  a  simple  example  of  the  applicability  of  the  multiscale  model, 
although  the  model  is  vary  general.  In  particular  the  state  at  any  node  can 
be  described  as  a  vector.  For  example,  one  may  represent  the  state  at  any 
node  by  a  vector  whose  elements  are  the  components  of  the  surface  energy 
balance  equation.  The  model  can  then  be  used  to  estimate  the  energy  balance 
at  various  scales  by  combining  point  and  remote  sensing  estimates.  These 
issues  are  currently  under  investigation. 


(7) 


(8) 


104 


Etimeled  Field 


10  20  30  40  SO  60 

Estimated  -  Observed  Field 


10  20  30  40  5  0  60  10  20  30  40  50  60 

Figure  4:  This  figure  shows  the  original  volumetric  soil  moisture  field  (top-left) 
and  the  estimated  field  using  a  fractal  prior  model.  The  lower  row  shows  the 
estimation  error  and  the  difference  field. 

REFERENCES 

1.  Avissar,  R.  and  R.  Pielke,  A  parameterization  of  Heterogeneous  Land- 
Suraface  for  Atmospheric  Numerical  Models  and  its  Impact  on  Regional 
Meteorology,  Mori.  Weather  Rev.,  117,  2113-2136,  1989. 

2.  Basseville,  M.,  A.  Benveniste,  and  A.  S.  Willsky,  Multiscale  Autoregres- 
sive  Processes,  Parts  1  and  2,  IEEE  Trans,  on  Signal  Processing,  40, 
1915-1954,  1992. 

3.  Chou,  K.  C.,  A.  S.  Willsky,  A  Benveniste,  Multiscale  Recursive  Estima- 
tion, Data  Fusion,  and  Regularization,  IEEE  Trans,  on  Auto.  Control, 
39(3),  1994. 

4.  Daley,  R.,  Atmospheric  Data  Analysis,  Cambridge  University  Press,  1991. 

5.  Mahrt,  L.,  Grid-Averaged  Surface  Fluxes,  Mon.  Weather  Rev.,  115,  1550- 
1560,  1987. 


105 


6.  Rodriguez-Iturbe,  I.,  G.  K.  Vogel,  R.  Rigon,  D.  Entekhabi,  F.  Castelli,  A. 
Rinaldo,  On  the  Spatial  Organization  of  Soil  Moisture  Fields,  Geophysical 
Res.  Letters,  22(20),  2757-2760,  1995. 

7.  Sellers,  P.  J.,  F.  G.  Hall,  G.  Asrar,  D.  E.  Strebel,  and  R.  E.  Murphy,  An 
Overview  of  the  First  International  Satellite  Land  Surface  Climatology 
Project  (ISLSCP)  Field  Experiments  (FIFE),  Jul.  of  Geophysical  Res., 
97(D17),  18345-18371,  Nov.,  1992. 

8.  Smith,  E.  A.  et  al.,  Area- Averaged  Surface  Fluxes  and  Their  Time-Space 
Variability  Over  the  FIFE  Experimental  Domain,  Jnl.  of  Geophysical  Re- 
search, 97(D17),  18599-18622,  1992. 

9.  Stofan,  E.  R.,  el  al.,  Overview  of  Results  of  Spaceborne  Imaging  Radar- 
C,  X  Band  Synthetic  Aperture  Radar  (SIR-C/X-SAR),  IEEE  Trans,  on 
Geoscience  and  Remote  Sensing,  33(4),  817-828,  1995. 

10.  Valencia,  D.,  and  J.  C.  Schaake,  Jr.,  Disaggregation  Processes  in  Stochas- 
tic Hydrology,  Water  Resources  Res.,  9(3),  580-585,  1973. 

11.  Wetzel,  P.  J.,  and  J.  T.  Chang,  Evapotranspiration  from  Nonuniform  Sur- 
faces: A  First  Approach  for  Short-Term  Numerical  Weather  Prediction, 
Mon.  Weather  Rev.,  116,  600-621,  1988. 

BIOGRAPHICAL  SKETCH 

Praveen  Kumar  is  an  Assistant  Professor  in  the  Department  of  Civil  En- 
gineering at  the  University  of  Illinois  in  Urbana-Champaign,  Illinois.  He  ob- 
tained his  PhD  from  the  University  of  Minnesota  in  1993.  His  research  interests 
include  analysis  and  modeling  of  multiscale  variability  in  surface  hydrologic 
processes,  large  scale  hydrologic  processes,  and  hydrometeorology  and  hydro- 
climatology. 


106 


Flow  Based  Scale-up  of 
Heterogeneous  Porous  Media 
using  Homogenization  and 
Wavelet  Representation 

Joe  Koebbe1  and  Ryan  Thomas2 


Abstract  -  Scaling  up  microscopic  heterogeneities  to  a  macroscopic 
level  useful  in  reservoir  engineering  applications  is  a  difficult  problem. 
Many  of  the  small  scale  heterogeneous  structures  such  as  shale  barriers 
that  occur  deltaic  rock  formations  dictate  flow  on  the  larger  scale.  Most 
averaging  methods  are  not  not  able  to  maintain  the  information  con- 
tained in  these  types  of  structures  at  the  macroscopic  scale.  Another 
important  example  is  that  of  a  fractured  reservoir  where  depending  on 
age,  fractures  can  act  as  either  conduits  or  barriers  to  flow  through  the 
rock.  The  goal  of  the  research  presented  is  to  use  (1)  flow  based  averag- 
ing techniques  to  construct  a  more  rigorous  way  of  averaging  the  rock 
properties  and  (2)  use  wavelet  representation  in  the  averaging  method 
to  preserve  the  effect  of  any  microscopic  structures  that  strongly  influ- 
ence the  flow.  Results  from  the  application  of  this  methodology  to  the 
rock  formations  at  the  Ferron  Sandstone  site  in  southern  Utah  will  be 
presented.  Transect  data  obtained  by  the  Utah  Geological  Survey  in 
joint  work  with  Mobil  Oil,  British  Petroleum,  and  a  number  of  other 
research  groups  will  be  used  as  a  basis  of  the  numerical  studies. 


SCALING  PROBLEMS  IN  POROUS  MEDIA 

Computer  modeling  of  fluid  flow  in  porous  media  is  used  to  predict  the 
performance  of  reservoirs  in  groundwater  flow  and  petroleum  engineering  ap- 
plications. In  many  cases  the  porous  medium  is  highly  heterogeneous.  The 
heterogeneity  occurs  at  all  scales  in  the  problem  and  very  small  scale  features 
of  a  reservoir  can  strongly  influence  flow  of  fluids  in  the  reservoir.  In  most 
applications,  computer  models  are  needed  on  a  scale  that  is  large  relative  to 
many  important  geological  features.   Exactly  resolving  the  important  small 

Associate  Professor,  Department  of  Math  &:  Stat,  Utah  State  University,  Logan,  UT 
2Undergraduate  Student,  Department  of  Math  &;  Stat,  Utah  State  University,  Logan, 
UT 


107 


scale  features  is  impossible  and  normally  some  method  of  averaging  is  used  in 
the  hope  that  the  bulk  or  averaged  properties  will  still  contain  the  important 
information  from  the  small  scale  heterogeneities. 

For  example,  fluvial  deltaic  formations,  which  comprise  a  large  share  of 
the  known  hydrocarbon  bearing  reservoir  formations  in  the  world,  are  created 
by  patterns  of  depositional  events  that  leave  possibly  large  deposits  of  coarse 
grain  material  such  as  sand,  separated  by  small  amounts  of  finer  grain  mate- 
rials, such  as  clays  and  silt.  The  resulting  rock  formation  from  these  types  of 
depositional  environments  is  layered  with  high  permeability  regions  separated 
by  thin,  relatively  low  permeability  layers.  The  bulk  flow  in  these  formations 
is  strongly  influenced  by  these  very  small  scale  features. 

Fractures  in  a  formation  are  also  small  scale  features  that  can  function  as 
either  conduits  or  barriers.  Just  after  formation  a  fracture  will  function  as 
a  conduit  allowing  fluid  to  move  from  one  region  to  another  easily.  Later  as 
minerals  deposit  along  the  fracture  surface  the  fracture  becomes  a  barrier  to 
flow.  Each  of  these  regimes  is  important  in  reservoir  modeling.  To  understand 
how  a  formation  has  trapped  the  hydrocarbons  it  necessary  to  be  able  to  model 
the  early  flow  history  through  the  fracture  and  to  understand  how  to  produce 
the  hydrocarbon  effectively  after  the  fractures  have  shut  off,  the  behavior  of 
the  fracture  as  a  barrier  to  flow  must  be  understood.  The  effects  of  these 
relatively  small  scale  heterogeneities  must  be  included  in  the  computer  model. 

AVERAGING  PROCESSES  AND  HOMOGENIZATION 

The  description  will  focus  on  a  simple  flow  model,  Darcy's  equation,  for  a 
porous  medium.  The  simplest  partial  differential  equation  that  models  flow  in 
a  reservoir  is 

V  •  K(x)VA(x)  =  /(*)  (1) 

The  parameters  in  the  problem  are  the  rock  permeability  (or  conductivity) 
tensor,  K  defined  by 

K  =  kxx,       K  — 

in  one  two  and  three  dimensions,  respectively,  the  pressure  head  (or  hydraulic 
head)  the  spatial  variable  x,  and  a  forcing  function  /  which  can  be  used 
to  represent  wells  or  confining  reservoirs.  For  simplicity  assume  that  the  per- 
meability varies  on  only  two  scales  which  will  be  termed  the  microscopic  and 
macroscopic  scales.  In  what  follows  the  only  parameter  that  will  be  scaled 
up  will  be  the  permeability,  K.  For  convenience  the  field  is  assumed  to  be 
periodic.  The  ideas  here  have  been  extended  to  aperiodic  permeability  fields. 


h  h 

h  h 
<^yx  ^yy 


h 

I^XX 

1c 

^xy 

h 

^xz 

3 

K  = 

h 

i^yx 

h 

h 

r^yz 

h 

h 

i^zy 

kZz 

108 


Reservoir 


X2 


o 

{J 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

oc 

o 

o 

o 

o 

o 

o 

oc 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

o 

— 

macroscopic  scale 


Y 


microscopic  scale 


Figure  1:  Illustration  of  the  (a)  macroscopic  scale  and  (b)  a  microscopic  view 
of  a  cell. 


Figure  (1)  depicts  the  two  length  scales  in  two  spatial  dimensions.  The 
pattern  of  the  permeabilities  from  a  distant  or  macroscopic  view  is  periodic.  As 
we  zoom  in  on  the  cells  we  see  that  in  each  cell  the  structure  of  the  permeability 
can  be  quite  general.  It  is  assumed  that  the  cell  is  small  compared  to  the 
overall  size  of  the  reservoir.  Next  define  two  length  scales;  the  macroscopic 
scale  represented  by  the  variable  x  and  the  microscopic  scale  represented  by  the 
variable  y.  The  macroscopic  scale  involves  the  pressure  and  velocity  variables 
over  the  entire  reservoir,  X.  The  microscopic  scale  is  defined  by  a  single  cell 
in  Figure  (1),  denoted  by  Y.  The  relationship  between  x  and  y  is  y  =  x / e. 
Note  that  that  on  the  microscopic  scale  as  y  varies  from  0  to  1  the  macroscopic 
variable  x  varies  from  0  to  e. 

The  goal  is  to  determine  a  bulk  permeability  on  the  region  X  by  scaling 
up  the  information  in  the  small  cells.  There  are  some  simple  cases  that  can  be 
discussed  without  complicated  averaging.  Consider  a  model  of  flow  through  a 
layered  porous  medium  with  well  defined  discontinuities  between  the  different 
layers  of  rock.  Assume  the  permeability  in  each  type  of  rock  is  constant.  If  we 
assume  that  the  fluid  flow  is  either  parallel  or  perpendicular  to  the  direction  of 
flow  then  the  scale  up  methods  that  are  appropriate  are  clear.  When  the  flow 
is  parallel  to  the  discontinuities  the  arithmetic  average  should  be  used  for  the 
permeability  and  when  the  flow  is  perpendicular  to  the  direction  of  flow  the 


109 


Permeability  =  10 


Permeability  =  1 


Permeability  =  10.0 


Permeabiility  =  1.0 


Figure  2:  The  symmetric  and  inverted-L  patterns  of  permeabilities  used  for 
illustrating  problems  with  functional  averaging  methods.  The  harmonic,  arith- 
metic and  geometric  means  return  diagonal  tensors. 


harmonic  average  of  the  permeabilities  should  be  used.  With  a  little  analysis 
it  is  easy  to  show  that  if  the  flow  occurs  at  some  other  angle  of  incidence 
to  the  discontinuities  the  permeability  should  be  between  the  harmonic  and 
arithmetic  average.  The  question  is  what  is  the  appropriate  effective  value  to 
used  based  on  the  flow?  In  many  cases  ad  hoc  methods  are  used  that  combine 
harmonic  and  arithmetic  averaging.  If  the  field  is  log-normal  the  geometric 
average  may  be  appropriate. 

Each  of  these  averaging  methods  is  a  functional  averaging  method.  That 
is,  no  matter  where  the  signal  or  data  has  come  from  the  same  averaging  tech- 
nique will  be  applied.  How  the  data  is  used  in  the  solution  of  the  flow  problem 
is  not  considered.  Thus  the  standard  methods  of  averaging  are  not  condi- 
tioned on  the  underlying  physical  process.  As  an  illustration  of  the  problem 
consider  the  two  permeability  patterns  shown  in  Figure  (2).  A  symmetric  and 
nonsymmetric  pattern  that  looks  like  an  inverted  'L'are  shown.  Each  of  these 
regions  can  be  divided  into  16  equally  sized  cells;  4  with  permeability  1  and 
12  with  permeability  10.  In  two  dimensions  the  permeability  tensors  in  this 
case  would  be 


K  = 


1  0 
0  1 


=  I 


or 


K  = 


10  0 
0  10 


=  10  1 


where  I  is  the  usual  2x2  identity  matrix.  Suppose  that  we  impose  a  pressure 
gradient  so  that  the  fluid  will  flow  from  lower  left  to  upper  right  in  the  region. 
We  should  expect  the  fluid  to  move  around  the  obstacle  in  both  case,  the  flow 
should  be  symmetric  in  the  first  case  while  asymmetric  in  the  inverted-L  case. 


110 


If  we  average  the  permeability  tensor  using  any  functional  based  method 
we  will  obtain  a  symmetric  tensor  which  is  a  multiple  of  the  identity.  The 
tensors  for  the  arithmetic,  harmonic,  and  geometric  averages  are  K a  =  7.75  I 
and  Kh  %  3.077  I,  and  Kg  ~  5.623  I,  respectively  for  both  the  symmetric  and 
asymmetric  patterns.  You  should  note  that  in  all  of  these  cases  the  principle 
flow  directions  are  still  lined  up  along  the  coordinate  axes.  Thus  using  these 
averaging  methods  will  produce  the  same  tensor  and  the  same  bulk  properties. 
We  need  an  averaging  method  that  is  conditioned  to  the  flow  problem  we  are 
trying  to  solve. 

The  homogenization  procedure  explained  below  will  generate  tensors 


K* 


6.651 
0 


0 

6.651 


an< 


K* 


5.574 
-0.348 


0.348 
6.949 


for  the  symmetric  and  asymmetric  patterns,  respectively.  For  the  symmetric 
pattern  the  principle  flow  directions  are  lined  up  along  the  coordinate  axes  as 
one  would  expect.  In  the  asymmetric  pattern  the  principle  flow  directions  are 
at  approximately  76  degrees  from  the  horizontal  axis. 

In  the  next  section  a  very  brief  description  of  the  homogenization  process 
will  be  given.  The  interested  reader  is  referred  to  Bourgeat,  1984,  for  the- 
oretical results  and  to  Amaziane,  et.  al.,  1990  for  computational  aspects  of 
homogenization.  In  addition,  a  code  exists  for  applying  homogenization  on 
rectangular  grids  Koebbe,  1996.  The  computer  code  can  be  downloaded  via 
anonymous  ftp  from 

sunf s . math . usu . edu/pub/koebbe/homog . tar . Z 

and  then  uncompressed  on  a  Unix  operating  system. 


The  Perturbation  Method 


The  averaging  is  done  over  one  of  the  small  cells  as  depicted  in  Figure  (1). 
In  the  work  below  we  will  want  to  use  the  Darcy  velocity  in  the  computations 
which  is  defined  by  v  =  — KV/i  This  quantity  can  be  used  to  reduce  the  second 
order  partial  differential  equation  to  a  system  of  first  order  partial  differential 
equations  of  the  form 

-V-v  =  /  (2) 
v   =    -KVh.  (3) 

The  perturbation  method  will  be  applied  to  the  system  of  partial  differential 
equations  defined  by  Equations  (2)  and  (3). 

The  parameter  e  which  is  related  to  the  size  of  the  microscopic  cell  is 
small;  that  is  0  <  e  <<  1.  The  perturbation  method  then  assumes  that  the 


111 


variables,  h  and  v,  can  be  expanded  in  a  series  in  the  perturbation  parameter  e; 

h  =  h0-\-6h1  +  e2h2-\  and  v  =  v0  +  ev!  +  e2 v2  H  The  spatial  differentiation 

operator  V  becomes  V  =  Vx  +  ^Vy  where  the  subscripts  denote  differentiation 
with  respect  to  to  the  specified  variable.  The  differentiation  is  decomposed  into 
two  pieces;  one  at  the  macroscopic  scale,  Vx,  and  one  at  the  microscopic  scale, 
Vy.  The  microscopic  derivative  is  scaled  by  ^  due  to  the  relationship  y  —  \ 
mentioned  above. 

The  multiple  scale  differentiation  operator  and  the  series  expansions  are 
then  substituted  into  the  first  order  system  in  Equations  (2)  and  (3).  The 
result  is  the  system 

-(V,  +  -Vy)(v0  +  cvi  +  •  •  •)   =   /  (4) 
vo  +  ev!  +  ..-    =    -K£(Vx+ivy)(/i0  +  e/il  +  ---)  (5) 

The  perturbation  method  equates  the  coefficients  multiplying  the  powers  of 
the  small  parameter  e  on  either  side  of  Equations  (4)  and  (5). 
The  equations  for  the  powers  of  e-1  and  e°  are: 

-V,  -v0  =  0  (6) 

0  =  KVyho  (7) 

and 

-(V..V0  +  Vy.Vl)  =  /  (8) 

-v0  =  K(Vxh0  +  Vj,/n),  (9) 

respectively.  The  coefficients  that  are  multiplied  by  a  power  of  e  greater  than 
zero  are  neglected.  From  this  point  on  the  perturbation  is  routine  and  the 
interested  reader  can  check  the  references  for  the  details  of  the  analysis. 

Computation  of  the  Homogenized  Permeability 

In  the  perturbation  analysis  it  is  necessary  to  assume  the  microscopic  pres- 
sure variable  h\  can  be  written  as  a  combination  of  the  partial  derivatives  of 
the  macroscopic  pressure  variable  ho  in  the  following  way. 

3  Qfo 
k=l  dxk 

The  multipliers  Wk{y)  allow  the  computation  of  a  scaled  combination  of  macro- 
scopic derivatives  and  can  be  thought  of  as  containing  the  first  order  fluctua- 
tions of  the  pressure  on  a  microscopic  cell,  Y. 


112 


With  this  assumption  the  perturbation  analysis  results  in  an  averaged  first 
order  system  of  equations 


_Vr.v#    =    /#  (n) 
v#   =   -K*Vh0         jV«  ¥f>  (12) 

These  averaged  or  macroscopic  equations  are  of  the  same  from  as  the  original 
equations  defined  for  the  original  heterogeneous  medium. 

The  effective  permeability,  K*,  returned  by  this  perturbation  analysis 


K*  =<  K(I  + 


is  define 

d  by 

dw\  dw2 

d  u>3  1 

dyi  dyi 

dyi 

dwi  d\U2 

3ti/3 

dy2  dy2 

dy2 

ami  ovj2 

dWQ 

dyz  dyz 

dyz  . 

)>Y  (13) 


where  the  notation  <  •  >  indicates  an  integral  average  over  the  domain.  The 
integral  average  is  done  for  each  of  the  nine  components  separately.  The 
problem  that  we  are  left  to  solve  is:  Find  the  functions  Wk(y)  that  are  included 
in  the  form  given  in  Equation  (13). 

If  we  substitute  the  form  in  Equation  (13)  into  Equation  (10)  we  generate 
a  system  of  uncoupled  elliptic  partial  differential  equations  of  the  form 

Vy  •  KVwk{y)  =  -VKe*  (14) 

where  k  =  1,2,3  and  e*.  is  the  kth  unit  vector  in  3£3  (or  3£  or  3£2  in  the  one 
and  two  dimensional  cases,  respectively).  The  solutions  of  these  equations  is 
done  via  some  standard  finite  element  method.  For  a  piecewise  constant  per- 
meability a  piecewise  continuous  linear  finite  element  will  return  exact  results 
and  also  the  continuity  of  flux  condition  is  a  natural  boundary  condition  for 
the  Galerkin  method. 

SPECTRAL  SOLUTIONS  AND  WAVELETS 

To  illustrate  the  connection  of  the  homogenization  process  to  wavelet  rep- 
resentation consider  the  one  dimensional  restriction  of  Equation  (14)  to  the 
unit  interval  [0, 1].  The  homogenized  permeability  is  computed  using 

K*=/K(l  +  ^  (15) 
Jo  ay 

where  the  function  w(y)  must  satisfy  the  ordinary  differential  equation 

— K—  =  -— K 

dy     dy  dy 


113 


Physically,  should  be  the  harmonic  average  and  it  is  always  the  case  that 
the  homogenization  process  agrees  with  this  value. 

The  problem  can  be  viewed  from  the  point  of  Sturm-Liouville  theory  if 
we  include  the  boundary  conditions.  The  conditions  that  are  imposed  by  the 
physical  problem  are  (1)  continuity  of  the  fluid  flux,  (2)  the  average  returned 
must  be  the  harmonic  average,  and  (3)  a  'wavelet'  condition  which  can  be 
stated  as  w(0).  Solving  Equation  (15)  with  the  given  conditions  will  produce 
a  set  of  orthogonal  eigenfunctions  (trigonometric  functions)  that  can  be  used 
as  a  basis  for  constructing  characterizations  of  the  permeability  field  from  the 
point  of  view  of  Fourier  analysis  or  wavelets.  Details  of  this  process  will  be 
presented  in  the  talk  and  appear  in  another  paper. 

REFERENCES 

Bourgeat,  A.,  1984  Homogenization  method  applied  to  the  behavior  of  a  natu- 
rally fissured  reservoir,  In  K.I.  Gross,  editor,  Mathematical  Methods  in 
Energy  Research,  pages  181-193.  SIAM. 

Amaziane,  B.,  Bourgeat,  A.,  and  Koebbe,  J.,  1990,  Numerical  simulation 
and  homogenization  of  two-phase  flow  in  heterogeneous  porous  media, 
Hornung,  Dogan,  and  Knaber,  editors,  Transport  in  Porous  Media  II. 
Kluwer  Academic  Publishers. 

Koebbe,  J.V.,  1996,  Homcode:  A  code  for  scaling  up  permeabilities  using 
homogenization,  (to  appear  as  a  Utah  Geological  Survey  report). 

Bourgeat,  A. P.,  and  Hidani  A.,  1994,  Effective  model  of  two-phase  flow  in  a 
porous  medium  of  different  rock  types.  Publication  de  l'Equipe  d'Analyse 
Numerique,  Lyon-St.  Etienne 

Bourgeat,  A. P.,  Kozlov,  S.M.,  and  Mikelic,  A.,  April  1993,  Effective  equations 
of  two-phase  in  random  media,  Publication  de  l'Equipe  d'Analyse 
Numerique,  Lyon-St.  Etienne 

BIOGRAPHICAL  SKETCH 

Joe  Koebbe  is  an  Associate  Professor  in  the  Department  of  Mathematics 
and  Statistics  at  Utah  State  University.  He  received  a  PhD  in  Mathematics 
from  the  University  of  Wyoming  in  1988.  Joe  works  in  reservoir  simulation 
with  applications  in  groundwater  flow  and  petroleum  applications. 

Ryan  Thomas  is  an  undergraduate  student  at  Utah  Stat  University  working 
towards  a  Bachelors  degree  in  mathematics  and  Masters  degree  in  electrical 
engineering.  His  area  of  interest  is  currently  in  signal  processing. 


114 


Nonhomogeneous 
Hidden  Markov  Models 
Allowing  Stochastic  Downscaling 
of  Synoptic  Atmospheric  Patterns 
to  Local  Hydrologic  Phenomena 

Peter  Guttorp1  and  James  P.  Hughes2 


Abstract. —  A  model  for  multistation  precipitation,  conditional  on  syn- 
optic atmospheric  patterns,  is  presented.  It  postulates  the  existence  of  an 
unobserved  weather  state  which  serves  as  a  link  between  the  large-scale 
atmospheric  measures  and  the  small-scale  spatially  discontinuous  precip- 
itation field.  The  weather  state  process  is  assumed  to  be  conditionally 
Markov,  given  the  atmospheric  data.  The  rainfall  process  is  then  condi- 
tionally specified  given  the  weather  state.  Various  parameterization s  for 
the  weather  state  process  and  the  rainfall  process  are  discussed  and  a 
likelihood  based  estimation  procedure  is  described.  As  an  example  the 
model  is  fit  to  a  twenty  station  network  of  rain  gauge  stations  in  western 
Washington.  We  discuss  the  NHMM  as  a  method  of  relating  synoptic 
atmospheric  data,  such  as  the  output  of  a  general  circulation  model,  to 
rainfall  and  other  local  (sub-grid  scale)  hydrologic  processes,  and  outline 
possible  extensions  to  continental  scales. 


INTRODUCTION 

Stochastic  downscaling  of  GCM  data  involves  building  a  stochastic  model 
to  relate  historical  measurements  of  synoptic  circulation  to  historical  regional  or 
local  climate  data  (e.g.  precipitation).  Then,  using  GCM  circulation  data  as  input, 
the  model  can  be  used  to  produce  simulations  at  the  regional  or  local  level.  The 
strength  of  this  procedure  lies  in  its  relative  simplicity  and  computational  ease 
compared  to  the  difficult  task  of  building  a  nested  GCM.  However,  the  procedure 
assumes  that  the  relationships  (between  circulation  and  regional  climate)  identi- 
fied in  the  historical  data  will  be  preserved  under  the  alternative  climate  scenarios 
produced  by  the  GCM. 

Initial  attempts  at  stochastic  downscaling  utilized  "weather  state"  models 
(e.g.  McCabe  et  al.  (1989),  Hay  et  al.  (1991),  Wilson  et  al.(1992),  Bardossy  and 
Plate  (1992)).  In  these  models,  each  day  is  explicitly  classified  into  one  of  a  few 


Department  of  Statistics,  Box  354322,  University  of  Washington.  Seattle.  WA  98195-4322. 
Department  of  Biostatistics,  University  of  Washington.  Seattle.  WA. 


115 


discrete  categories  (weather  states)  based  on  the  synoptic  circulation  pattern  for 
that  day.  Then  separate  rainfall  models  are  fit  conditional  on  the  weather  state. 
Hughes  et  al.  (1993)  developed  such  a  model  using  historical  circulation  and  rain- 
fall and  then  applied  the  results  to  data  from  a  GCM  2  x  C02  run  to  describe  rain- 
fall patterns,  streamflow  and  flood  frequency  under  the  altered  climate. 

A  difficulty  with  the  weather  state  models  as  defined  above  is  that  the 
weather  states  must  be  explicitly  defined  a  priori.  A  poor  choice  of  the  weather 
states  will  lead  to  a  poor  model.  As  an  alternative,  Zucchini  and  Guttorp  (1991) 
proposed  a  hidden  Markov  model  (HMM)  as  a  model  for  precipitation.  The  HMM 
presupposes  the  existence  of  a  few  discrete  weather  states  but  it  assumes  that 
these  states  cannot  be  observed  directly.  Instead,  only  rainfall  is  observed  and  the 
rainfall  is  assumed  to  be  conditionally  temporally  independent  given  the  weather 
state.  In  this  model  the  weather  states  are  effectively  defined  by  the  rainfall  pro- 
cess. The  states  chosen  are  those  that  separate  the  rainfall  process  into  different 
spatial  or  temporal  patterns.  Formally,  let  Rf  be  the  measurement  (typically  multi- 
variate) of  the  regional  or  local  process  (e.g.  rainfall  occurrence)  at  time  t  and  St 
be  the  weather  state  at  time  t.  Then  the  hidden  Markov  model  is  defined  by  the 
assumptions 

P(Rfl5[,R,r,)  =  P(Rrl5/)  (1) 

P(St\S\-l)  =  P(St\St_{)  (2) 

where  the  notation  S[  means  all  values  of  Sr  from  time  1  to  T.  The  term  hidden 
refers  to  the  fact  that  the  weather  state  process,  Sti  is  an  unobserved  quantity  and 
is  not  defined  a  priori  (although  the  weather  state  for  each  day  may  be  identified 
after  the  model  is  fit).  Nonetheless,  Zucchini  and  Guttorp  (1991)  found  that  the 
circulation  patterns  associated  with  particular  weather  states  were  interpretable.  In 
effect,  the  model  acts  as  an  automatic  classifier  of  circulation  patterns  into 
weather  states  that  are  associated  with  particular  precipitation  patterns. 

The  major  drawback  of  the  HMM  approach  is  that  it  fails  to  incorporate 
synoptic  circulation  information  in  the  definition  of  the  weather  states.  The 
weather  states  are  defined  solely  on  the  basis  of  rainfall.  Thus,  the  model  cannot 
be  used  for  downscaling.  However,  an  extension  to  the  HMM  which  does  incorpo- 
rate atmospheric  information  and  can  be  used  for  downscaling  was  developed  by 
Hughes  and  Guttorp  (1994a).  They  termed  their  model  a  nonhomogeneous  hid- 
den Markov  model  (NHMM).  Letting  Xr  be  the  measurement  (or  summary)  of 
the  atmospheric  data  at  time  t,  the  NHMM  is  defined  by  the  assumptions 

P(R,ISlr,R^X[)  =  P(R,IS,)  (3) 

P(Sr\S[-l,XTl)  =  P(St\St_l,Xt)  (4) 

The  first  assumption  says  that  the  local  or  regional  process,  Rr,  is  conditionally 
temporally  independent  given  the  current  weather  state.  In  other  words,  the 
regional  process  is  driven  by  the  weather  state  process.  The  second  assumption 


116 


states  that  the  (unobserved)  weather  state  process  is  Markov  with  transition  proba- 
bilities depending  on  the  current  values  of  the  atmospheric  information.  Various 
parameterizations  are  possible  for  P(5,I5,_| ,  X,)  and  P(R,\St)  and  maximum  like- 
lihood techniques  are  available  to  estimate  the  parameters.  Hughes  and  Guttorp 
i  lc)94a)  show  that  both  the  explicit  weather  state  models  mentioned  previously 
and  the  hidden  Markov  model  are  special  cases  of  the  more  general  NHMM. 

Hughes  and  Guttorp  (1994a)  applied  the  NHMM  to  24  years  of  winter  data 
from  a  four  station  rain  gauge  network  in  Washington  state  using  the  following 
model  for  P(R,\St): 

P(R,  =  rt\St  =  s)  =  ft,  piO  ~  P,y~r''  (5) 

This  model  assumes  that  the  probability  of  rain  at  each  station  depends  only  on 
the  weather  state  and  is  independent  of  the  other  stations.  For  the  network  ana- 
lyzed by  Hughes  and  Guttorp  (1994a)  this  assumption  was  reasonable  since  the 
rain  stations  were  widely  separated.  Note  that  such  an  assumption  does  not  imply 
that  the  rainfall  occurrence  processes  at  the  stations  are  uncorrelated.  Hughes  and 
Guttorp  (1994a)  found  strong  (unconditional)  correlations,  on  the  order  of  0.4  to 
0.5,  between  the  stations.  This  correlation  is  induced  by  the  common  weather 
state. 

In  Hughes  and  Guttorp  (1994b),  the  NHMM  was  applied  to  21  years  of 
winter  data  in  a  densely  packed  network  of  24  stations  in  the  Puget  Sound  region 
of  western  Washington  state.  The  assumption  of  spatial  independence  (even  con- 
ditional on  the  weather  state)  was  no  longer  tenable.  Therefore,  the  following 
model  was  used  for  P(Rr\St): 

P(Rr  =  rr\St  =  s)  -  exp  {yU  ^/  +  E;<«  ftrjr//^  ]  (6) 

V  J 

where  di}  is  the  distance  between  stations  i  and  j.  The  parameters  ps  measure  the 
spatial  dependence  between  stations.  When  ps  is  positive,  stations  i  and  j  are 
positively  correlated  (within  weather  states)  and  the  correlation  decreases 
inversely  with  distance.  A  negative  value  for  J3S  implies  negative  correlation 
between  stations  i  and  j  (within  weather  states).  When  each  ps  is  0,  equation  (6) 
reduces  to  the  independence  model  (equation  (5))  with  als  =  \og{plsl{  \  -  pls)).  As 
before,  maximum  likelihood  may  be  used  for  parameter  estimation.  However,  for 
large  numbers  of  stations  the  constant  of  proportionality  in  equation  (6)  is  difficult 
to  evaluate,  and  Hughes  and  Guttorp  (1994b)  describe  an  alternative  estimation 
strategy  for  that  situation. 

When  applied  to  the  24  station  network  mentioned  above,  and  based  on 
mean  sea-level  pressure  and  the  north-south  gradient  of  500  mb  geopotential 
height,  the  model  identified  five  weather  states.  State  1  had  a  southwesterly  flow 
over  the  Puget  Sound  region,  both  at  the  surface  and  aloft,  resulting  in  high  pre- 
cipitation probabilities  everywhere.  State  2  had  westerly  surface  flow,  but  south- 
west flow  aloft.  There  was  substantial  rainfall  in  the  north  and  south,  but  relatively 


117 


low  probability  of  rain  in  the  central  part  of  the  region.  State  3  was  characterized 
by  southwesterly  flow  at  sea  level  and  westerly  flow  aloft,  with  fairly  weak  sur- 
face winds.  The  stations  bordering  Puget  Sound  were  in  a  rain  shadow,  while 
more  easterly  stations  had  high  precipitation  probabilities.  State  4  was  character- 
ized by  a  high  pressure  system  over  the  region,  with  consequent  low  precipitation 
rates.  Finally,  state  5  was  similar  to  state  2,  but  with  a  more  westerly  500  mb  flow. 
The  rainfall  pattern  was  similar  to  that  in  state  2. 

Statistics  of  the  historical  rainfall  process  (i.e.  probability  of  rain,  spatial 
and  temporal  correlations,  storm  duration  distribution)  were  generally  well  repro- 
duced although  stations  at  the  periphery  of  the  network  or  isolated  stations  tended 
to  exhibit  a  predicted  (simulated)  duration  distribution  which  was  lighter-tailed 
than  that  which  was  actually  observed. 

MODELING  PRECIPITATION  ON  LARGER  SCALES 

The  NHMM  has  not  yet  been  applied  on  subcontinental  or  continental 
scales.  Some  modification  of  the  model  will  be  necessary  in  this  case  since,  in  the 
present  configuration,  St  is  a  finite  scalar  quantity  which  is  meant  to  describe  the 
weather  state  at  time  t  over  a  (meteorologically)  homogeneous  region.  One  possi- 
bility would  be  to  identify  K  distinct  precipitation/meteorological  regions  and 
then  allow  St  to  be  a  A'-variate  vector  where  the  k-th  element  corresponds  to  the 
weather  state  in  the  k-ih  region.  The  model  for  P(St\St_\,  Xt)  therefore  becomes  a 
model  for  a  multivariate  Markov  process.  The  simplest  such  model  is 

P(St\St_uXt)  =  UP(Sit\Sit_l,Xt)  (7) 

which  postulates  that  each  regional  process  evolves  independently,  conditional  on 
the  observed  atmospheric  process,  Xt.  Presumably,  one  would  also  want  to 
assume  that  the  rain  stations  in  each  region  depend  only  on  that  region's  hidden 

state: 

P(Rt\St)  =  YlP(rtU\Slt)  (8) 

where  R\1'  denotes  the  set  of  rain  stations  in  region  i.  Note  that  this  assumption  of 
spatial  independence  between  regions  does  not  preclude  the  use  of  a  model  for 
spatial  dependence  within  each  region. 

Unfortunately,  it  seems  unlikely  that  the  situation  will  be  as  simple  as  the 
(conditional)  regional  independence  model  envisioned  in  (1).  It  is  possible  that 
some  enhancement  in  skill  can  be  gained  by  coupling  weather  states  in  each 
region  with  weather  states  in  other  regions.  For  instance,  the  movement  of  large 
scale  storm  fronts  across  regions  would  suggest  that  yesterday's  weather  state  in 
one  region  could  provide  information  about  today's  weather  state  in  another 
region.  On  the  other  hand,  if  N  =  Ylk=i  dim(S  )  is  the  total  number  of  states  in  the 
AT-variate  process  St,  it  is  not  computationally  feasible  to  attempt  to  fit  the  entire 
N(N-\)  transition  matrix,  P(S,\St_\tXt).  A  reasonable  compromise  between 
these  extremes  might  be  to  assume  that  today's  state  in  region  i  depends  only  on 


118 


yesterday's  state  in  the  regions  surrounding  region  i.  That  is, 

P(S,\St_l,Xl)  =  YlP(Sit\S^,Xt)  (9) 

r 

where  fAjO  is  read  "the  neighborhood  of  i".  Basic  meteorologic  principles  and 
knowledge  of  the  regions  under  consideration  should  be  used  to  determine  which 
regions  constitute  the  neighborhood  of  region  i. 

Another  approach  to  modeling  continental  scale  precipitation  would  be  to 
divide  the  atmospheric  variables  into  region-specific  measures  (call  these  Xt '  for 
region  /)  and  (sub)continental  measures  (call  these  Xt  )  and  to  postulate  the  exis- 
tence of  an  unobserved  continental  scale  weather  state  (call  this  Ct  ).  Under  these 
assumptions,  one  might  write 

P(St\St_ltXt)  =  Ec,  P(ct\x?ni  P(Slt\Sll9Xf'9Ct)  (10) 

In  other  words,  the  weather  state  in  region  i  today  depends  on  the  weather  state  in 
region  i  yesterday,  the  atmospheric  data  in  region  i  today  and  the  overall  continen- 
tal weather  state  today.  The  probability  distribution  of  the  continental  weather 
state  is  completely  defined  by  the  continental  atmospheric  variables. 

SUMMARY 

Simulations  of  precipitation  conditional  on  a  set  of  atmospheric  data  can  be 
used  to  downscale  the  output  of  a  GCM  into  local  precipitation,  and  to  assess  the 
regional  and  local  effects  of  climate  change  scenarios  (by  conditioning  the  simula- 
tions on  the  output  of  GCM  runs  assuming  altered  climate  scenarios).  This  does, 
of  course,  require  the  untestable  assumption  that  precipitation  reacts  to  the  altered 
climate  in  the  same  fashion  as  it  does  to  the  historical  climate.  In  other  words,  we 
must  assume  that  the  climate  change  affects  only  the  distribution  of  weather 
states,  rather  than  which  weather  states  are  appropriate. 

REFERENCES 

Bardossy,  .A.  and  E.  J.  Plate  (1992)  Space-time  models  for  daily  rainfall  using 
atmospheric  circulation  patterns.  Water  Resour.  Res.,  28,  1247-1259. 

Hay,  L.,  G.  J.  McCabe,  D.  M.  Wolock,  and  M.  A.  Ayers  (1991)  Simulation  of  pre- 
cipitation by  weather  type  analysis.  Water  Resour.  Res.,  27,  493-501. 

Hughes,  J.  R,  D.  R  Lettenmaier,  R  Guttorp  (1993)  A  stochastic  approach  for 
assessing  the  effects  of  changes  in  regional  circulation  patterns  on  local  pre- 
cipitation. Water  Resour.  Res.,  29,  3303-3315. 

Hughes,  J.  R  and  R  Guttorp  (1994a)  A  Class  of  Stochastic  Models  for  Relating 
Synoptic  Atmospheric  Patterns  to  Regional  Hydrologic  Phenomena.  Water 
Resour.  Res.,  30,  1535-1546. 

Hughes,  J.  P.  and  P.  Guttorp  (1994b)  Incorporating  spatial  dependence  and  atmo- 
spheric data  in  a  model  of  precipitation.  J.  Applied  Meteor.,  30,  1535-1546. 


119 


McCabe,  G.  J.,  L.  E.  Hay,  M.  A.  Ayers,  and  D.  M.  Wolock  (1989)  Assessment  of 
climate  change  using  weather  type  analysis.  Presented  at  the  National  Con- 
ference on  Hydraulic  Engineering,  Am.  Soc.  of  Civ.  Eng.,  New  Orleans, 
La.,  Aug.  14-18. 

Wilson,  L.  L.,  D.  P.  Lettenmaier  and  E.  Skyllingstad  (1992)  A  hierarchical 
stochastic  model  of  large-scale  atmospheric  circulation  patterns  and  multi- 
ple station  daily  precipitation.  J.  Geophys.  Res.,  97.  2791-2809. 

Zucchini,  W.  and  P.  Guttorp  (1991)  A  hidden  Markov  model  for  space-time  pre- 
cipitation. Water  Resour.  Res.,  27,  1917-1923. 


120 


Explicit  Consideration  of  Multiple 
Landscape  Scales  While  Selecting 
Spatial  Resolutions 

John  B.  Collins  1  and  Curtis  E.  Woodcock  2 


Abstract.  -  The  information  content  of  a  digital  image  -  defined  as 
the  spatial  variation  of  the  image  data  -  depends  on  data  resolution. 
Spatial  data  may  vary  at  a  number  of  different  scales  simultaneously, 
so  the  choice  of  an  appropriate  resolution  depends  on  the  information 
content  at  the  landscape  scale  relevant  to  a  particular  application.  A 
useful  framework  for  studying  variation  at  different  scales  in  digital 
images  is  provided  by  a  nested  hierarchical  model  of  the  landscape. 
The  landscape  is  modeled  as  being  composed  of  spatial  units  of  a  typi- 
cal size,  which  are  themselves  composed  of  smaller  units,  and  so  on. 
The  overall  variation  of  the  image  data  can  be  decomposed  into  varia- 
tion due  to  effects  of  different  levels  of  such  a  hierarchy.  The  hierarchi- 
cal decomposition  of  the  variation  in  a  digital  image  is  reflected  by  a 
similar  hierarchical  decomposition  of  the  scene  semi  variance.  Decom- 
posed semivariograms  are  used  to  assess  the  relationships  between  data 
resolution  and  information  content  at  particular  landscape  scales. 


INTRODUCTION 

The  issue  of  scale  is  of  fundamental  importance  to  disciplines  concerned 
with  phenomena  which  vary  spatially.  Yet  the  relationship  between  the  scale  at 
which  a  phenomenon  is  observed  and  that  phenomenon's  characteristic  scale  of 
variation  is  not  completely  understood.  In  this  discussion,  scale  of  observation 
is  defined  as  the  spatial  interval  over  which  observations  are  made.  When  data 
are  collected  by  remote  sensing  devices,  this  term  corresponds  to  sensor  resolu- 
tion, or  size  of  the  Instantaneous  Field  of  View  (IFOV).  The  characteristic 
scale  of  a  phenomenon  is  the  range  of  resolutions  over  which  that  phenomenon 
may  be  characterized.  These  definitions  correspond  to  those  proposed  by  Quat- 
trochi  (1993). 


Research  Assistant,  Boston  University  Department  of  Geography.  Boston,  MA. 
Associate  Professor  of  Geography,  Boston  University.  Boston,  MA. 


121 


In  remote  sensing  investigations,  resolution  is  an  important  consideration. 
The  information  contained  in  an  image,  which  can  be  identified  as  the  overall 
variance  of  the  image  data,  declines  as  resolution  becomes  more  coarse.  The 
simplest  view  of  the  issue  is  one  which  regards  the  selection  of  an  appropriate 
resolution  as  a  compromise  between  greater  information  content  (fine  resolu- 
tions) and  more  manageable  data  volume  (coarse  resolutions).  But  for  some 
applications,  greater  spatial  variation  is  not  always  desireable.  For  example, 
classification  of  remotely  sensed  imagery  is  most  successful  when  pixel  size  is 
large  enough  that  some  amount  of  averaging  of  within-class  variation  occurs 
(Markham  and  Townshend,  1981). 

The  fact  that  variation  of  spatial  data  occurs  at  multiple  scales  has  an 
explanation  in  the  field  of  landscape  ecology.  It  is  becoming  an  accepted  tenet 
of  that  field  that  landscapes  are  often  arranged  hierarchically  (O'Neill  et  al., 
1986).  That  is,  landscapes  are  often  composed  of  objects  which  are  internally 
homogeneous  and  externally  variable  with  respect  to  spatially  varying 
processes.  These  objects  may  be  composed  of  smaller  objects,  and  so  on.  For 
example,  a  landscape  may  be  composed  of  a  number  of  watersheds,  each  con- 
taining a  collection  of  vegetation  stands,  which  each  contain  individual  plants. 
Associated  with  each  level  of  such  a  hierarchy  is  a  set  of  processes  acting  at 
that  level.  The  range  of  sizes  of  objects  on  a  given  level  approximately  defines 
the  spatial  frequencies  at  which  those  processes  vary  -  i.e.  their  characteristic 
scales.  So  multiple  scales  of  variation  of  spatial  data  arise  from  a  scene's 
hierarchical  structure. 

The  choice  of  an  appropriate  resolution  for  a  particular  investigation 
depends  on  prior  knowledge  of  the  characteristic  scale  of  variation  of  the 
phenomenon  of  interest.  This  paper  presents  methods  to  help  make  such  a 
choice  under  explicit  consideration  of  the  hierarchical  nature  of  landscapes.  A 
review  of  the  use  of  hierarchical  partitionings  of  spatial  data  is  presented,  fol- 
lowed by  an  explanation  of  geostatistical  methods  for  investigating  resolution- 
dependent  characteristics  of  digital  images. 

SPATIAL  HIERARCHIES 

A  number  of  studies  use  a  hierarchical  model  to  identify  multiple 
landscape  scales.  A  common  paradigm  involves  partitioning  a  scene  into 
objects  of  regular  size  and  shape  -  usually  squares  or  rectangles  -  each  of 
which  are  characterized  by  some  aggregation  of  the  underlying  spatially  distri- 
buted variable  (Grieg-Smith,  1963;  Townshend  and  Justice,  1995).  Then  there 
is  a  clear  and  explicit  relationship  between  any  level  of  the  hierarchy  and  a 
scale  of  observation.  By  examining  some  attribute  of  each  level  of  the  hierar- 
chy -  especially  the  data  variance  -  multiple  scales  of  variation  may  be 
revealed. 

An  alternate  approach  is  the  use  of  hierarchies  of  irregular-shaped  objects, 
such  as  political  boundaries  (Moellering  and  Tobler,  1972).  The  characteristic 
scales  associated  with  any  level  are  identified  with  the  range  of  sizes  of  objects 
on  that  level.  However,  the  hierarchies  so  constructed  can  be  rather  arbitrary. 


122 


Woodcock  et  al.  (1994)  describe  a  forest  mapping  project  which  makes  use  of  a 
hierarchical  model  dividing  the  landscape  into  ecologically  meaningful  units, 
(see  also  Woodcock  and  Harward,  1992;  Collins  et  al.,  1995).  Under  this 
model,  the  entire  image  is  considered  to  be  composed  of  a  number  of  vegeta- 
tion classes  such  as  conifer,  hardwood,  and  brush.  Each  of  these  can  be  bro- 
ken down  into  smaller  homogeneous  patches,  referred  to  as  regions.  Within 
the  conifer  class,  for  example,  a  region  corresponds  to  a  forest  stand.  Regions, 
in  turn,  are  collections  of  individual  plants.  Such  a  hierarchy  can  be  associated 
with  ecosystem  processes,  which  tend  to  be  more  homogeneous  within  units 
and  heterogeneous  between  units  at  any  level. 

The  partitioning  of  a  scene  into  ecologically  meaningful  units  prior  to 
analysis  constitutes  explicit  identification  of  the  characteristic  scales  under 
study.  Knowledge  of  these  characteristic  scales  is  important  for  assessing  the 
relationship  between  resolution  and  information  content  for  a  given  application. 

Statistical  Analysis  of  Hierarchal  Data 

A  general  and  mathematically  precise  definition  of  a  hierarchical  scene 
model  is  needed.  Consider  a  regionalized  variable  Z  which  depends  on  vector 
location  s  within  some  domain  set  D/0).  When  discussing  analysis  of  digital 
images,  D(0)  is  a  subset  of  two-dimensional  space  R2,  and  Z  could  be  radiance 
in  some  wavelength  band  or  any  other  spatially-varying  quantity.  The  hierarch- 
ical model  consists  of  partitioning  the  domain  set  into  a  family  of  objects 
(sets)  {£>(!)},  which  form  level  1  of  a  hierarchy.  Each  of  the  level-1  objects 
may  be  divided  into  a  number  of  level-2  objects,  which  can  collectively  be 
referred  to  as  {D^2)}-  Further  partitions  proceed  in  an  obvious  manner.  If  level 
n  is  the  highest  level  of  the  hierarchy,  then  the  family  of  domain  sets  {£(„)}  is 
the  collection  of  infitesimal  points  on  which  Z  is  defined.  Any  point  s  e  D(0) 
will  be  contained  in  exactly  one  set  associated  with  each  level  of  the  hierarchy. 
Such  a  set  of  level  k  is  D^k^(s).  This  terminology  is  illustrated  by  Figure  1. 


Figure  1.  A  hierarchical  partitioning  of  a  scene.  A.  The  set  of  level-1  objects  {D^} 
forming  a  partition  of  D^Qy  B.  The  level-2  objects  {D^)}  further  partitioning  the  level-1 
objects. 


123 


Determination  of  the  relative  contributions  of  different  levels  of  a  hierar- 
chy to  overall  variation  in  a  data  set  is  the  goal  of  Hierarchical  Analysis  of 
Variance  (Moellering  and  Tobler,  1972).  Under  the  Hierarchical  ANOVA 
model,  the  observed  value  at  a  point  in  the  scene  is  considered  equal  to  the 
overall  mean  for  the  area,  plus  a  deviation,  or  effect,  characteristic  of  each  level 
of  the  hierarchy.  The  effect  of  level  k  of  a  hierarchy  at  point  s  is  defined  as 
D(k)(s  )~D(k-\)(s),  where  D^(s)  is  the  mean  value  of  Z  over  the  level-&  object 
containing  s .  Note  that  each  effect  constitutes  a  regionalized  variable  in  its 
own  right.  Further  elaboration  of  the  Hierarchical  ANOVA  model  can  be 
found  in  standard  texts  on  the  subject  (e.g.  Dunn  and  Clark,  1974).  The  model 
leads  to  the  following  important  result. 

=  °2(i)  +  <*2(2)  +  '  '  '  +°2(n)    .  (1) 

where  a2z  represents  the  overall  variance  of  Z,  and  a2(i)  represents  the  vari- 
ance of  the  effect  of  level  i.  The  overall  data  variance  decomposes  into  the 
sum  of  variances  of  each  effect.  Unfortunately,  the  ability  to  make  statistical 
inferences  from  a  hierarchical  partitioning  of  variance  relies  on  the  assumption 
of  independence  of  observations.  Since  this  assumption  is  badly  violated  for 
most  spatial  data  sets,  the  technique  is  suited  only  for  descriptive  purposes. 

When  the  regionalized  variable  Z  exhibits  appropriate  stationarity  charac- 
teristics, its  variation  is  described  by  the  scene  semivariogram  (Isaaks  and 
Srivastava,  1989;  Cressie,  1993). 

yz(h)=±E[Z(s)-Z(s+h)]  ,  (2) 

where  h  is  a  vector  lag.  Similarly,  relationships  between  two  regionalized  vari- 
ables Y  and  Z  are  described  by  the  cross-semivariogram. 

JYz(h)  =  ^E{[Z(s)-Z(s+h)][Y(s)-Y{s+h)]}  .  (3) 

The  parameters  describing  the  scene  semivariogram  -  its  range,  sill,  and 
derivative  at  the  origin  -  can  be  shown  to  be  related  to  quantifiable  attributes 
of  a  scene  (Woodcock  et  al.  1988a+b). 

Since  the  semivariogram  provides  a  better  characterization  of  the  variation 
of  a  regionalized  variable  than  does  the  variance,  it  is  useful  to  ask  whether 
there  is  a  geostatistical  analogy  to  Eq.  1.  Collins  et  al.  (1995)  present  an 
example  to  show  that  under  the  Hierarchical  ANOVA  model,  the  scene 
semivariogram  decomposes  as 

yz(h)  =  y(l)(h)  +  •••  +y(n)(fc)  +  2  isYo-x*)^)  .  (4) 

;=1*=1 

where  yz  represents  the  overall  scene  semivariogram,  y(/)  represents  the 
semivariogram  associated  with  the  effect  of  level  i  of  the  hierarchy,  and  Y(/)(*) 
is  the  cross-semivariogram  between  the  effects  of  levels  j  and  k .  It  is  interest- 
ing to  note  that  when  overall  scene  variance  is  considered  (Eq.  1)  all  cross- 
product  terms  vanish,  but  that  this  is  not  the  case  for  spatial  variance.  The 
cross-semivariograms  in  (4)  arise  as  the  result  of  fine-scale  continuity  across 


124 


boundaries  of  larger-scale  objects  (Collins  and  Woodcock,  manuscript  in 
preparation).  They  generally  have  small  magnitudes  and  relatively  short 
ranges,  and  approach  zero  at  large  lags.  The  relevance  of  Eq.  4  to  the  present 
study  is  its  indication  that  semivariograms  associated  with  particular  levels  of  a 
hierarchy  -  i.e.  with  different  characteristic  scales  in  the  landscape  -  may  be 
discussed  individually. 


EFFECT  OF  RESOLUTION  ON  SPATIAL  STRUCTURE 


The  scale-specific  semivariograms  included  in  Eq.  4  may  be  used  to 
retrieve  quantitative  information  about  scene  variation  at  arbitrary  resolutions. 
In  geostatistical  parlance,  sensor  resolution  -  the  size  and  shape  of  the  IFOV  - 
is  called  the  support  of  the  measurements.  There  exists  a  well-developed 
theory  of  the  effect  of  changes  of  support  on  regionalized  variables  (Rendu. 
1978;  Zhang  et  al.,  1990),  and  its  applications  in  the  context  of  remote  sensing 
(Jupp  et  al.,  1988+1989;  Atkinson  and  Curran,  1995).  This  section  discusses 
some  of  the  ways  in  which  information  relating  to  scale  of  observation  can  be 
derived  from  semivariograms.  Notationally,  subscripts  indicative  of  levels  in  a 
hierarchy  are  dropped,  as  the  ideas  apply  to  any  regionalized  variable  with 
finite  variance.  The  relevance  to  multiple  scales  of  variation  comes  from  the 
indication  of  Eq.  4  that  these  methods  can  be  applied  to  scale-specific 
semivariograms. 


Regularization  and  Deregularization 


As  was  noted  above,  the  effect  of  a  level  in  a  hierarchy  is  a  regionalized 
variable,  and  is  defined  at  all  points  in  a  scene.  However,  point-scale  observa- 
tions are  never  available  from  a  remote  sensing  device,  since  there  is  always 
some  averaging  of  scene  radiance  over  the  sensor  IFOV.  That  is,  the  data  and 
the  semivariograms  derived  from  it  are  regularized.  It  is  desireable  to  derive 
the  point-scale,  or  punctual  semivariograms  from  the  regularized  ones. 

In  the  theoretical  development  which  follows,  the  support  of  observations 
is  modeled  as  a  set  in  the  domain  space  D(0)  over  which  the  continuously- 
varying  values  of  some  effect  are  averaged.  A  useful  quantity  is  the  average 
punctual  semivariance  between  two  sets  in  the  scene,  denoted  A  and  B . 

t**B)=  1      —J  J  y{z-z)dz'  dz  (5) 

Mes(A  )Mes(# )  A  B 

Mes(*)  denotes  the  Lebesque  measure  of  a  set  (i.e.  its  area  in  this  context). 
This  definition  is  particularly  useful  when  the  sets  A  and  B  are  shifts  of  a  sen- 
sor IFOV.  Specifically,  let  w  denote  some  support,  and  let  vv/2  denote  its  trans- 
lation by  distance  h  .  The  regularized  semivariogram  at  a  lag  of  h  can  be 
found  from  the  punctual  semivariogram  by  (Rendu,  1978) 

Yw  (h)  =  7(w  Wh )  -  Y(vv  ,w )  ,  (6) 
This  relationship  cannot  be  inverted  to  find  the  punctual  semivariogram  from 


125 


its  regularized  counterpart.  But  Atkinson  and  Curran  (1995)  present  a  numeri- 
cal technique  based  on  this  relationship  which  may  be  used  to  estimate  it. 

The  effects  of  regularization  are  most  pronounced  on  higher  levels  of  a 
scene  hierarchy,  where  the  characteristic  scales  of  variation  are  the  smallest. 
When  scales  of  variability  are  much  smaller  than  the  support,  'microstructures' 
in  the  data  are  undetected  and  spatial  variation  is  lost.  This  phenomenon  plus 
the  effect  of  measurement  error  can  possibly  lead  to  an  observed  'nugget 
effect'  or  a  non-zero  intercept  of  a  regularized  semivariogram.  At  scales  larger 
than  that  at  which  data  are  collected,  the  nugget  effect  shows  no  spatial  struc- 
ture. This  leads  to  a  simple  relationship  between  the  nugget  effect  observed 
via  support  w  (Nw)  and  that  observed  at  a  larger  support  W  (Nw): 

Mes(w)  AT 

Nw  =  —LNW   .  (7) 

w     Mes(W)    w  v  ; 

The  nugget  effect  behaves  differently  under  changes  of  support  than  does 
the  spatially  correlated  component  of  variation.  Thus  it  is  useful  to  express  the 
regularized  semivariogram  as  (Zhang  et  al.,  1990): 

yw(h)=Nw  +yw(h)  ,  (8) 

where  y*^  is  a  regularized  zero-nugget  semivariogram,  from  which  a  punctual 
zero-nugget  semivariogram  y  may  be  estimated.  Then  using  6,  7,  and  8  the 
semivariogram  for  any  support  W  can  be  expressed  as 

Jw(h)  =  Nw  +Y(W,Wh) -Y(W.W)  .  (9) 

Since  parameters  describing  semivariograms  relate  to  properties  of  the  scene 
under  study,  the  ability  to  determine  semivariograms  for  arbitrary  supports  is 
useful.  When  used  with  a  hierarchical  decomposition  of  a  scene 
semivariogram,  it  can  yield  useful  information  concerning  the  relationship 
between  image  resolution  and  spatial  variation  at  different  characteristic  scales. 

Determination  of  Variance  at  Arbitrary  Resolutions 

As  was  mentioned  above,  the  overall  scene  variance  is  perhaps  the  best 
measure  of  its  information  content.  As  a  rule,  the  variance  decreases  as  the 
size  of  support  increases.  But  the  precise  rate  of  decrease  depends  on  the  auto- 
correlation structure  of  the  image,  as  revealed  by  the  semivariogram.  If  auto- 
correlation between  points  is  strong,  then  regularization  will  have  little  effect 
since  averaging  over  the  IFOV  will  tend  to  occur  between  points  with  similar 
values. 

This  assertion  is  justified  mathematically.  Let  D(0)  represent  an  area  being 
studied.  If  this  area  has  been  sampled  on  supports  w ,  then  the  variance  of  sup- 
ports W  within  D(0)  is  (Zhang  et  al.,  1990) 

Mes(w)  Mes(w) 


V(W,D(0))  = 


Mes(W)  Mes(D(0)) 


Nw  +y  (Z)(0),D(0))  -  y  (W,W)  (10) 


126 


The  first  term  accounts  for  the  nugget  effect  among  the  supports  W,  and  the 
last  two  terms  deal  with  the  spatially-correlated  component  of  variation.  Eq. 
10  gives  the  variance  for  arbitrary  resolutions,  and  so  is  useful  for  a  number  of 
purposes. 

CONCLUSION 

Methods  to  determine  the  effect  of  sensor  resolution  on  different 
landscape  scales  have  been  presented.  Characteristic  scales  of  variation  can  be 
identified  by  partitioning  a  scene  hierarchically  into  ecologically  meaningful 
units.  In  order  not  to  restrict  the  discussion  to  a  particular  classification  of 
scene  objects,  a  general  hierarchical  model  has  been  used.  Use  of  different 
landscape  hierarchies  may  be  called  for  in  different  situations,  but  the  ideas  dis- 
cussed here  continue  to  apply. 

Semivariograms  may  be  calculated  characterizing  the  spatial  variation  at 
each  landscape  scale,  and  may  be  used  to  derive  information  about  scene  struc- 
ture at  arbitrary  resolutions.  Throughout  the  discussion  above,  the  assumption 
has  been  made  that  the  data  are  second-order  stationary  at  all  scales,  a  require- 
ment for  the  calculation  of  semivariograms.  This  assumption  needs  to  be 
justified  in  order  for  the  analysis  of  semivariance  to  be  appropriate.  If  this  con- 
dition is  met,  the  relationships  between  resolution  and  semivariance  presented 
here  are  general  with  respect  to  directional  qualities  of  variation.  All  results 
apply  equally  for  isotropic  and  non-isotropic  semivariograms. 

REFERENCES 

Atkinson,  P.  M.  and  Curran,  P.  J.  (1995).  Defining  an  optimal  size  of  support 
for  remote  sensing  investigations,  IEEE  Transactions  on  Geoscience  and 
Remote  Sensing,  33(3):768-776. 

Collins,  J.  B.,  Woodcock,  C.  E.  and  Jupp,  D.  L.  B.  (1995).  Spatial  dependence 
and   nested   hierarchical    scene   models,    1995  ACSM/ASPRS  Annual 
Convention  and  Exposition  Technical  Papers: 
535-544. 

Cressie,  N.  A.  C.  (1993).  Statistics  for  Spatial  Data,  John  Wiley  and  Sons, 
Inc.,  New  York. 

Dunn,  O.  J.  and  Clark,  V.  A.  (1974).  Applied  Statistics:  Analysis  of  Variance 
and  Regression,  John  Wiley  and  Sons,  Inc.,  New  York. 

Hobbs,  R.  J.  and  Mooney,  H.  A.,  Eds.  (1990).  Remote  Sensing  of  Biosphere 
Functioning,  Springer- Verlag,  New  York. 

Isaaks,  E.  H.  and  Srivastava,  R.  M.  (1989).  An  Introduction  to  Applied 
Geostatistics,  Oxford  University  Press,  Oxford. 

Jupp,  D.  L.  B.,  Strahler,  A.  H.,  and  Woodcock,  C.  E.  (1988).  Autocorrelation 
and  regularization  in  digital  images  I:  basic  theory,  IEEE  Transactions  on 
Geoscience  and  Remote  Sensing,  26(4):463-473. 


127 


Jupp,  D.  L.  B.,  Strahler,  A.  H.,  and  Woodcock,  C.  E.  (1989).  Autocorrelation 
and  regularization  in  digital  images  II:  simple  image  models,  IEEE 
Transactions  on  Geoscience  and  Remote  Sensing,  27(3):247-258. 

Markham,  B.  L.,  and  Townshend,  J.  R.  G.  (1981).  Land  cover  classification 
accuracy  as  a  function  of  sensor  spatial  resolution,  Proceedings  of  the  15th 
Int.  Symp.  on  Remote  Sensing  of  Environment,  Ann  Arbor,  Mi.:  1075- 
1090. 

Matheron,  G.  (1963).  Principles  of  geostatistics,  Economic  Geology,  58:1246- 
1266. 

Mollering,  H.  and  W.  Tobler  (1972).  Geographical  variances,  Geographical 
Analysis,  4  (1): 35-50. 

O'Niell,  R.  V.,  DeAngelis,  D.  L.,  Waide,  J.  B.,  and  Allen,  T.  F.  H.  (1986).  A 
Hierarchical  Concept  of  Ecosystems,  Princeton  University  Press, 
Princeton,  New  Jersey. 

Quattrochi,  D.  A.  (1993).  The  need  for  a  lexicon  of  scale  terms  in  integrating 
remote  sensing  data  with  geographic  information  systems,  Journal  of 
Geography,  92(3):206-212. 

Rendu,  J.  M.  (1978).  An  Introduction  to  Geostatistical  Methods  of  Mineral 
Evaluation,  South  African  Institute  of  Mining  and  Metallurgy, 
Johannesburg. 

Townshend,  J.  R.  G.  and  Justice,  C.  O.  (1995).  Spatial  variability  of  images 
and  the  monitoring  of  changes  in  the  normalized  difference  vegetation 
index,  International  Journal  of  Remote  Sensing,  16(12):21 87-2195. 

Woodcock,  C.E.,  Collins,  J.,  Gopal,  S.,  Jakabhazy,  V.,  Li,  X.,  Macomber,  S., 
Ryherd,  S.,  Wu,  Y.,  Harward,  V.  J.,  Levitan,  J.,  and  Warbington,  R. 
(1994).  Mapping  Forest  Vegetation  Using  Landsat  TM  Imagery  and  a 
Canopy  Reflectance  Model,  Remote  Sensing  of  Environment,  50:240-254. 

Woodcock,  J.  E.  and  J.  Harward  (1992).  Nested-Hierarchical  Scene  Models 
and  Image  Segmentation,  International  Journal  of  Remote  Sensing, 
13(16):3167-3187. 

Woodcock,  C.  E.,  Strahler,  A.  H.,  and  Jupp,  D.  L.  B.  (1988).  The  use  of 
variograms  in  remote  sensing:  I.  scene  models  and  simulated  images, 
Remote  Sensing  of  Environment,  25:323-348. 

Woodcock,  C.  E.,  Strahler,  A.  H.,  and  Jupp,  D.  L.  B.  (1988).  The  use  of 
variograms  in  remote  sensing:  II.  real  digital  images,  Remote  Sensing  of 
Environment,  25:349-379. 

Zhang,  R.,  Warrick,  A.  W.,  and  Meyers,  D.  E.  (1990).  Variance  as  a  function 
of  sample  support  size,  Mathematical  Geology  22(1):  107-121. 


128 


Using  Trend  Surface  Methodology 
to  Compare  Spatial  Surfaces 

Dale  Zimmerman1,  Zhi-Jun  Liu2,  and  George  Hallberg3 

Trend  surface  analysis  is  a  very  popular  method  for  describing 
spatial  data.  When  data  on  more  than  one  response  variable  are 
available,  it  is  generally  desirable  to  compare  the  correspond- 
ing fitted  surfaces  obtained  by  trend  surface  analysis.  Existing 
methods  for  such  comparisons  have  serious  deficiencies.  The  ob- 
jective of  this  article  is  to  develop  statistically  valid  techniques 
for  such  comparisons  which  retain  the  familiarity  and  relative 
simplicity  of  trend  surface  analysis.  In  particular,  we  develop 
tests  for  surface  parallelism  and  surface  coincidence.  The  meth- 
ods are  illustrated  by  applying  them  to  groundwater  quality  data 
from  an  aquifer  in  northeast  Iowa.  

INTRODUCTION 

The  multiple  linear  regression  method  known  as  polynomial  trend  surface 
analysis  (TSA)  is  used  widely  in  the  physical  sciences  to  describe  geostatistical 
data.  TSA  is  based  on  an  assumption  that  the  available  data  can  be  usefully 
described  by  a  model  in  which  each  observation  is  equal  to  the  sum  of  a 
trend  component,  generally  taken  to  be  a  polynomial  function  of  the  spatial 
coordinates,  and  an  error  component.  For  example,  for  data  taken  from  a 
two-dimensional  region,  a  pth-order  TSA  model  is  given  by 

Z(x,  y)  =  Poo  +  fox  +  fay  +  (32ox2  +  Pnxy  +  P02y2  +  •  *  *  +  PrSxrys  +  e(x,  y). 

Here,  Z(x,  y)  is  the  variate  of  interest's  value  at  the  site  with  Cartesian  coor- 
dinates (x,  y);  fioo,  Pio,  •  •  • ,  Prs  are  unknown  parameters;  r  +  s  =  p,  the  order  of 
the  polynomial;  and  e(x,  y)  is  the  error  component.  In  the  earliest  applications 
of  TSA,  inference  on  model  parameters  was  based  on  an  ordinary  least  squares 
fit  and  thus  the  errors  were  taken,  either  explicitly  or  implicitly,  to  be  inde- 
pendent normally  distributed  random  variables  with  zero  mean  and  constant 
variance  (Grant  1957;  Krumbein  1959).  Standard  F-tests  from  multiple  re- 
gression analysis  were  used  to  determine  the  order  of  the  trend  surface.  Later, 
however,  it  was  argued  that  in  many  situations  small-scale  positive  spatial 
dependence  is  likely  to  exist  among  the  errors,  rendering  the  ordinary  least 

Associate  Professor  of  Statistics,  University  of  Iowa,  Iowa  City,  IA. 
2 Research  Associate,  University  Hygienic  Laboratory,  Iowa  City,  IA. 
3 Geologist  and  Associate  Director,  University  Hygienic  Laboratory,  Iowa  City,  IA. 


129 


squares  estimates  of  model  coefficients  inefficient,  biasing  tests  on  the  param- 
eters, and  tending  to  result  in  a  fitted  surface  of  too  high  an  order  (Cliff  and 
Ord  1973;  Ripley  1981).  More  recently,  Agterberg  (1984)  and  Haining  (1987) 
have  used  generalized  least  squares  to  fit  TSA  models  with  spatially  correlated 
errors. 

Although  some  geostatistical  problems  in  which  TSA  can  be  used  involve 
fitting  a  surface  to  data  on  a  single  response  variable,  others  involve  fitting 
surfaces  to  data  on  several  response  variables.  If  these  variables  are  commen- 
surate or  correspond  to  a  single  variable  measured  on  several  occasions,  it  may 
be  of  interest  to  compare  the  surfaces  to  each  other.  For  example,  one  may 
want  to  compare  the  surface  that  describes  current  levels  of  a  groundwater 
contaminant  in  an  aquifer  to  the  surface  that  corresponds  to  levels  determined 
10  years  previously.  The  objective  of  this  paper  is  to  extend  the  statistical 
methodology  of  TSA  to  handle  such  comparisons. 

The  comparison  of  two  or  more  fitted  polynomial  trend  surfaces  is  not  a 
new  problem;  indeed,  it  was  an  active  area  of  research  in  the  1950s  and  1960s  as 
TSA  was  becoming  popular  among  mathematical  geologists  and  geographers. 
Three  main  approaches  to  the  problem  were  proposed:  a  point  correlation 
approach,  a  regression  coefficient  comparison,  and  a  gradient-based  approach. 
For  a  review  of  these,  see  Goodman  (1983).  Although  these  approaches  are 
popular,  they  have  several  shortcomings,  not  the  least  of  which  is  that  none 
of  them  yield  a  statistically  valid  test  for  surface  similarity  (Zimmerman  et  al. 
1996). 

The  methodology  for  comparing  spatial  surfaces  that  we  propose  here  does 
not  suffer  from  these  shortcomings.  In  particular,  it  permits  one  to  formally 
and  validly  test  for  coincidence  and  parallelism  of  the  surfaces  in  a  manner 
which  accounts  for  spatial  correlation. 

METHODOLOGY 
Preliminaries 

To  simplify  the  exposition,  we  consider  an  important  particular  case  of  the 
general  problem.  We  assume  that  only  two  surfaces,  with  common  spatial 
domain  D,  are  to  be  compared.  The  surfaces  may  correspond  either  to  two 
commensurate  variables  or  to  the  same  variable  measured  on  two  occasions; 
the  terminology  we  use,  however,  presumes  the  latter.  We  also  assume  that 
D  C  R2  and  denote  an  arbitrary  location  in  D  as  s  =  (x,y)'.  The  cases  of 
more  than  two  surfaces  or  more  than  two  dimensions  can  be  handled  by  the 
same  general  methods  to  be  presented  here,  but  the  notation  becomes  overly 
cumbersome. 

The  basis  of  our  approach  for  comparing  surfaces  is  to  regard  the  data  from 
which  each  fitted  surface  is  constructed  as  a  partial  realization  of  a  random 
process.  Let  {Zi(s):  s  £  D}  and  {Z2(s):  s  6  D}  denote  the  random  processes 


130 


corresponding  to  the  variable  of  interest  (or  to  perhaps  a  transformation,  such 
as  logarithms,  of  the  variable  of  interest)  on  the  two  occasions.  It  is  assumed 
that  these  processes  are  governed  by  the  model 


Zfc(s)  =    £    PijkxV  +  ek{s) 


(1) 


\<i+j<p 


where  {/^j*:  0  <  i  +  j  <  p;  k  =  1, 2}  are  unknown  parameters  and  {ei(s)}  and 
{62(s)}  are  independent,  stationary,  zero-mean  Gaussian  processes  with  covari- 
ance  functions  C\(h)  =  cov{ei(s),  ei(s  +  h)}  and  C2(h)  =  cov{e2(s),  e2(s  +  h)}, 
respectively.  We  assume  initially  that  p  and  the  covariance  functions  are 
known;  subsequently  we  will  discuss  their  estimation. 

Let  Si  =  (x\,yi)\  sni  =  (xni,yni)'  denote  the  n\  locations  where 
{Zi(s)}  is  observed,  and  let  sni+1  =  (xni+i,  yni+i)',  sni+U2  =  (xni+n2, 
yni+n2)'  denote  the  n2  locations  where  {Z2(s)}  is  observed.  Regarding  the 
relationship  of  the  first  set  of  locations  to  the  second  set,  two  cases  will  be 
considered:  (1)  arbitrary  sampling  networks,  in  which  no  relationship  between 
the  first  set  and  second  set  is  specified,  and  (2)  identical  sampling  networks, 
in  which  n\  =  n2  =  n  and  Sj  =  sn+/  for  I  =  1, . . . ,  n  (perhaps  after  a  suitable 
permutation  of  the  locations'  labels).  The  second  case  permits  an  analysis 
based  on  paired  differences  among  observations  at  the  same  locations  and 
thus  bears  a  resemblance  to  the  case  of  paired-difference  data  in  a  classical 
two-sample  problem.  If,  for  fixed  s,  the  errors  ei(s)  and  e2(s)  actually  exhibit 
positive  temporal  dependence  rather  than  being  independent  as  assumed  in 
model  (1),  then  the  variance  of  the  difference  in  these  two  errors  will  be  smaller 
than  the  sum  of  their  two  variances.  This  reduction  in  variation,  if  large  enough 
and  consistent  across  locations,  can  result  in  substantially  more  powerful  tests 
for  surface  coincidence  and  parallelism. 

Finally,  let  Zx  =  [Zi(sn), . . . ,  Zi(sini)]'  and  Z2  =  [Z2  (s2i ),...,  Z2(s2n2)]' 
denote  the  vectors  of  responses  observed  at  the  two  times,  and  let  ei  = 
[ei(sn),. .  .,€i(sini)]'  and  e2  =  [e2  (s2i )....,  e2(s2n2)]'  denote  the  correspond- 
ing vectors  of  errors.  Put  Z  =  [Z'1?  Z2]'  and  e  =  [e\,  e'2}'. 

Arbitrary  Sampling  Networks 

Under  the  stated  assumptions  and  using  the  notation  defined  previously, 
the  model  for  the  observed  data  can  be  written  as 


where  E(e)  =  0  and  var(e)  =  diag(En,  S22)  =  S,  which  we  take  to  be 
positive  definite.  Here,  T,kk  is  the  nk  x  nk  matrix  with  (r,  s)th  element  given 
by  Ck(skr  —  sks).  Let  q  =  (p+  l)(p  +  2)/2  and  let  Xi  denote  the  (ni  +  n2)  x  2q 
model  matrix  corresponding  to  model  (2),  which  is  assumed  to  have  full  column 


for  / 
for  / 


n\  -f  1, . .  .,rai  +n2 


(2) 


131 


rank,  and  let  (3  denote  the  associated  2q  x  1  vector  of  coefficients.  The  best 
linear  unbiased  estimator  of  f3  is  given  by  ft  =  (X.['E~lXi)~1X.'1Yi~1Z. 

For  purposes  of  testing  for  parallelism  or  coincidence  of  the  two  surfaces  in 
model  (2),  two  submodels  are  defined  as  follows. 
Parallelism  Model 

z  _  |  Pooi  +  Ei<i+j<P  PijX]y{  +  ei  for  /  =  1, . . . ,  m 

\  P002  +  Ei<i+j<p  PijXlyj  +ei   for  /  =  nx  +  1, . . . ,  nx  +  n2 

Coincidence  Model 

Zi  =  Poo  +  +  c/  for  /  =  1, . . . ,  ni  +  n2. 

i<*+j<p 

Now  let  #SSm  =  Z'E^Z  -  Z,E.-1Xm(X^E-1Xm)-1X^S-1Z  for  m  = 
1,2,3,  where  X2  and  X3  are  the  model  matrices  corresponding  to  the  paral- 
lelism and  coincidence  models,  respectively.  Then,  the  parallelism  hypothesis 
can  be  tested  by  comparing  the  value  of 

=  (RSS2  -  RSSl)/(g  -  1) 
P  RSSl/{nl+n2-2q) 

to  a  suitable  percentage  point  of  a  (central)  F  distribution  with  q  —  1  and 
ni  +  n2  —  2q  degrees  of  freedom.  Similarly,  we  can  test  for  coincidence  by 
comparing  the  value  of 

{RSSs-RSSJ/q 
C  RSSl/(nl+n2-2q) 

to  a  suitable  percentage  point  of  an  F  distribution  with  q  and  ri\  +  n2  —  2q 
degrees  of  freedom. 

The  procedures  just  described  are  identical  to  the  classical  tests  for  par- 
allelism or  coincidence  of  regression  surfaces  described,  for  example,  by  Weis- 
berg  (1985)  except  that,  to  account  for  spatial  correlation,  the  residual  sums 
of  squares  used  here  are  based  on  generalized  least  squares  (GLS),  rather  than 
ordinary  least  squares  (OLS),  fits  of  the  models.  It  should  be  noted  that  a 
"spatial  analysis  of  variance,"  of  which  the  analyses  of  variance  that  corre- 
spond to  these  procedures  are  examples,  has  been  used  before  for  problems 
other  than  testing  for  parallelism  or  coincidence  of  surfaces;  see  Griffith  (1978) 
and  Gotway  and  Cressie  (1990). 

Identical  Sampling  Networks 

Now  suppose  that  ri\  —  n2  =  n  and  =  sn+;  for  /  =  1, . . . ,  n.  Define  the 
paired  differences 

Dt   =   Z2(st)  -  Zi(si)  =  Zn+i  -  Zh 
Si   =   e2(sz)  -  ei(sj)  =  en+i  -  et  for  I  =  1, . . . ,  n 


132 


and  let  D  =  (Dt)  and  6  =  (Si).  It  follows  from  model  (2)  that  the  model  for 
D\  is 

A  =  loo  +  liMy{  +  ffi  for  /  =  1, . . .  ,n  (3) 

l<t+7<P 

where  7tJ  =  %2  -  ftji,  £(<5)  =  0.  and  var(<5)  =  En  +  £22  =  E*.  Let  W 
denote  the  n  x  q  model  matrix  corresponding  to  model  (3)  and  let  7  denote 
the  associated  q  x  1  vector  of  coefficients.  Based  on  this  model,  the  best  linear 
unbiased  estimator  of  7  is  given  by  7  =  (W'E^1W)_1  W'E^D. 

Testing  for  parallelism  corresponds,  in  this  situation,  to  testing  that  all  the 
coefficients  in  model  (3)  except  700  equal  0  and  hence  can  be  tested  by  the 
mean-corrected  overall  F-test  from  a  GLS  analysis.  That  is,  we  compare  the 
value  of 

RegSS,/(q-l) 
P'       RSS4/{n  -  q) 

to  a  suitable  percentage  point  of  an  F  distribution  with  q  —  1  and  n  — 
q  degrees  of  freedom,  where  RegSSA  =  D'S^1W(W,S71W)-1W,E^ JD  - 
D'E^l^E^l^-^E^D  and  RSS,  =  D'E^D-D'E^WtW'E^W)-1 
xW'Sj  D.  Testing  for  coincidence  corresponds  in  this  situation  to  testing 
that  all  coefficients  in  model  (3)  equal  0  and  hence  can  be  tested  by  comparing 

RegSS5/q 
C'  ~  RSS4/{n  -  q) 

to  a  suitable  percentage  point  of  an  F  distribution  with  q  and  n  —  q  degrees 
of  freedom,  where  RegSS5  =  D'E^WtW'E^W^W'E^D. 

Covariance  Estimation  and  Order  Determination 

In  practice,  of  course,  the  covariance  matrices  Sn  and  S22  (or  if  a 
difference-based  approach  is  possible)  and  the  orders  pi  and  p2  of  the  two  sur- 
faces are  unknown.  The  covariance  matrices  can  be  estimated  using  a  classical 
geostatistical  approach  based  on  the  sample  semivariogram;  see  Cressie  (1991. 
ch.  2).  Once  £  (or  5]$)  has  been  estimated,  the  tests  for  parallelism  and 
coincidence  that  we  have  described  can  be  carried  out  by  substituting  S  for 
E  in  the  previously  given  expressions  for  the  test  statistics.  Of  course,  the 
resulting  test  statistics  no  longer  have  F  distributions,  but  provided  that  S  is 
a  consistent  estimator  of  E,  the  asymptotic  distributions  of  the  test  statistics 
are  F  distributions  and  hence  the  tests  are  asymptotically  valid;  see  Cressie 
(1991.  pp.  242-243). 

As  for  the  determination  of  the  order(s)  of  the  surfaces,  a  stepwise  proce- 
dure is  described  in  Zimmerman  et  al.  (1996)  in  which,  starting  with  p0  =  0. 
a  (po  +  l)-order  model  is  fitted  by  OLS:  its  residuals  are  tested  for  spatial  cor- 
relation: and  then  an  F-test  (based  on  either  an  OLS  or  GLS  fit)  comparing 
the  p0-°rder  and  (p0  +  l)-order  models  is  carried  out. 


133 


EXAMPLE:  NITRATE  CONCENTRATION  IN 
BIG  SPRING  BASIN,  IOWA 

Big  Spring  Basin  is  a  groundwater  basin  located  in  Clayton  County  in 
northeast  Iowa.  It  is  named  after  Big  Spring,  the  largest  spring  in  the  state, 
which  discharges  from  the  underlying  Galena  aquifer.  The  area  of  the  basin 
is  267  square  km  (103  square  miles).  Because  of  the  proximity  of  the  aquifer 
to  the  land  surface  and  the  existence  of  karst  features  (such  as  sinkholes)  at 
the  surface,  the  Galena  aquifer  in  this  basin  is  highly  responsive  to  recharge 
events  and  susceptible  to  a  relatively  rapid  influx  of  contaminants.  The  land 
in  the  basin  is  essentially  all  used  for  agriculture.  There  are  no  significant 
urban  or  industrial  areas,  landfills,  or  other  major  point  sources  that  may 
affect  groundwater  quality.  The  combination  of  pure  agricultural  use  and  a 
responsive  hydrogeological  system  in  the  basin  offers  a  good  opportunity  for 
studying  the  impact  of  agricultural  non-point  source  pollution  on  groundwater 
quality. 

The  Geological  Survey  Bureau  of  the  Iowa  Department  of  Natural  Re- 
sources conducted  a  basin-wide  sampling  of  private  wells  in  the  Galena  aquifer 
in  1989,  which  was  a  drought  year,  and  1992,  which  was  a  relatively  wet  year. 
Wells  were  selected  somewhat  haphazardly,  and  inclusion  in  the  sample  de- 
pended on  the  presence  and  cooperation  of  the  property  owner,  among  other 
factors.  Complete  data  were  available  from  both  years  for  125  wells.  All  wells 
were  sampled  during  a  period  of  7-10  days  in  October  of  the  respective  year 
so  that  the  effects  of  short-term  climatic  and  hydrologic  events,  as  well  as  sea- 
sonal variability,  would  be  minimized.  Concentrations  of  several  contaminants 
in  the  samples  were  analyzed  by  the  University  Hygienic  Laboratory  in  Iowa 
City,  a  U.S.  Environmental  Protection  Agency  certified  contract  laboratory. 
Here  we  consider  only  the  nitrate  (NO3  )  concentrations,  which  were  measured 
in  mg/L  and  transformed  to  the  log  scale.  In  addition  to  the  nitrate  concen- 
trations, the  depth  and  (x,  y)  coordinates  of  each  well  was  recorded.  All  three 
coordinates  were  scaled  to  lie  in  the  range  [—1,1]. 

One  of  the  objectives  of  the  investigation  was  to  determine  whether  the  sur- 
face of  nitrate  concentrations  over  the  basin  was  consistent  with  what  would  be 
expected  from  scientific  considerations.  For  example,  it  is  well  known  that  ni- 
trate concentrations  decrease  with  increasing  depth.  Moreover,  in  the  eastern 
half  of  the  basin  there  were  a  large  number  of  sinkholes,  so  higher  concen- 
trations were  expected  there.  Another  objective  of  the  investigation  was  to 
determine  how  the  nitrate  concentration  surface  had  changed  from  1989  to 
1992.  In  particular,  since  precipitation  levels  differed  substantially  in  the  two 
years,  it  was  hoped  that  changes  in  the  surface  could  lead  to  the  formation 
of  tentative  hypotheses  on  how  precipitation  amounts  and  responsivity  of  the 
basin  may  interact  to  affect  groundwater  nitrate  concentrations. 

To  address  these  objectives,  first  we  performed  a  trend  surface  analysis 


134 


separately  for  each  year.  Application  of  our  order  determination  procedure 
resulted  in  a  first-order  model  for  both  1989  and  1992.  Examination  of  the 
sample  semivariograms  of  residuals,  as  well  as  Schmoyer's  (1994)  permutation 
test  for  spatial  correlation,  indicated  that  the  spatial  correlation  was  quite 
weak  at  the  scale  represented  by  most  lags  in  the  data  set.  Consequently,  we 
fit  first-order  models  by  OLS,  obtaining  the  fitted  models 

Zl(d,x,y)=  1.66  +  2.22a:  -  0.79w  -  lAbd 

(0.27)  (0.55)        (o.42)  (0-54) 

for  1989  and 

Z2(d,x,y)  =  2.52  +  1.90a;  -  0.45?/  -  1.72d 

(0.31)  (0.63)        (0.48)  (0-61) 

for  1992.  (Here,  d  represents  depth  and  quantities  in  parentheses  under  the 
coefficients  indicate  their  standard  errors.)  The  signs  and  magnitudes  of  the 
estimated  coefficients  were  consistent  with  what  was  expected  from  scientific 
considerations:  (1)  nitrate  concentration  decreased  with  increasing  depth;  (2) 
nitrate  concentration  increased  as  one  moved  from  west  to  east;  and  (3)  the 
relationship  between  nitrate  concentration  and  the  north-south  coordinate  was 
not  as  strong  as  the  others. 

Next,  we  compared  the  surfaces  by  testing  for  parallelism  and  coincidence 
using  the  F-tests  we  have  previously  described.  The  parallelism  hypothesis 
was  accepted  (P  =  .95)  but  the  coincidence  hypothesis,  assuming  parallelism, 
was  strongly  rejected  (P  <  .0001).  This  indicates  that  differences  in  precipi- 
tation levels  (and  possibly  differences  in  other  factors  confounded  with  year) 
significantly  affected  the  height,  but  not  the  shape,  of  the  nitrate  concentration 
surface. 

The  entire  testing  procedure  was  repeated  using  GLS  fits  of  the  two  sur- 
faces. The  classical  geostatistical  procedure  was  used  to  estimate  the  covari- 
ance  matrices  for  these  fits.  Results  of  the  tests  were  qualitatively  very  similar 
to  the  results  obtained  under  the  assumption  of  no  spatial  correlation. 

One  interesting  feature  of  the  OLS  analysis,  however,  was  that  the  residuals 
from  1989  were  significantly,  positively  correlated  with  the  residuals  (at  the 
same  sites)  from  1992  (r  =  0.70,  P  <  .0001).  This  suggests  that  although 
it  may  be  reasonable  to  assume  that  En  and  £22  are  scalar  multiples  of  an 
identity  matrix  for  these  data,  it  may  not  be  reasonable  to  assume  that  X)  is 
block  diagonal,  as  in  model  (2).  Furthermore,  there  was  evidence  of  somewhat 
greater  variation  among  the  measurements  in  1992  than  in  1989.  For  these 
reasons,  and  since  the  data  lie  on  identical  sampling  networks,  we  carried  out 
the  difference-based  approach  as  well.  The  conclusions  drawn  from  applying 
this  approach  were  not  fundamentally  different  than  those  obtained  using  the 
original  (i.e.  undifferenced)  observations. 


135 


REFERENCES 

Agterberg,  F.P.  1984.  Trend  surface  analysis.  Pages  147-171  in  Spatial  Statis- 
tics and  Models,  eds.  G.L.  Gaile  and  C.J.  Willmott,  Dordrecht:  D.  Reidel. 

Cliff,  A.  and  Ord,  J.  1973.  Spatial  Autocorrelation.  London:  Pion. 

Cressie,  N.  1991.  Statistics  for  Spatial  Data.  New  York:  Wiley. 

Gotway,  C.A.  and  Cressie,  N.A.C.  1990.  A  spatial  analysis  of  variance  applied 
to  soil-water  infiltration.  Water  Resources  Research  26,  2695-2703. 

Goodman,  A.  1983.  COMPARE:  A  Fortran  IV  program  for  the  quantitative 
comparison  of  polynomial  trend  surfaces.  Computers  &  Geosciences  9,  417- 
423. 

Grant,  F.A.  1957.  A  problem  in  the  analysis  of  geophysical  data.  Geophysics 
22,  309-344. 

Griffith,  D.A.  1978.  A  spatially  adjusted  ANOVA  model.  Geographical  Anal- 
ysis 10,  296-301. 

Haining,  R.  1987.  Trend-surface  models  with  regional  and  local  scales  of  vari- 
ation with  an  application  to  aerial  survey  data.  Technometrics  29,  461-469. 

Krumbein,  W.C.  1959.  Trend  surface  analysis  of  contour- type  maps  with  ir- 
regular control-point  spacing.  Journal  of  Geophysical  Research  68,  5869- 
5878. 

Ripley,  B.D.  1981.  Spatial  Statistics.  New  York:  Wiley. 

Schmoyer,  R.L.  1994.  Permutation  tests  for  correlation  in  regression  errors. 
Journal  of  the  American  Statistical  Association  89,  1507-1516. 

Weisberg,  S.  1985.  Applied  Linear  Regression.  New  York:  Wiley. 

Zimmerman,  D.L.,  Liu,  Z.  and  Hallberg,  G.  1996.  Using  trend  surface  method- 
ology and  locally  weighted  regression  to  compare  spatial  surfaces.  Un- 
published Technical  Report,  Department  of  Statistics  &  Actuarial  Science, 
University  of  Iowa. 

BIOGRAPHICAL  SKETCH 

Dale  Zimmerman  is  an  Associate  Professor  in  the  Department  of  Statistics 
and  Actuarial  Science  at  the  University  of  Iowa.  He  graduated  from  Iowa 
State  University  in  1986  with  a  Ph.D.  in  Statistics.  Dale  teaches  and  conducts 
research  in  spatial  statistics,  environmental  statistics,  and  linear  models. 

Zhi-Jun  Liu  is  a  Research  Associate  in  the  University  Hygienic  Laboratory 
at  the  University  of  Iowa.  He  received  an  M.S.  in  Statistics  and  a  Ph.D.  in 
Geography  from  the  University  of  Iowa.  His  research  interests  include  envi- 
ronmental modeling,  spatial  statistics,  and  geographic  information  systems. 

George  Hallberg  is  Associate  Director  of  the  University  of  Iowa  Hygienic 
Laboratory  and  an  Adjunct  Professor  in  Geology  and  Civil  and  Environmental 
Engineering.  He  received  his  Ph.D.  in  Geology  from  the  University  of  Iowa 
in  1975.  Hallberg  conducts  research  in  environmental  monitoring  and  public 
health  surveillance. 


136 


The  Effect  of  Spatial  Covariance 
Heterogeneity  on  Prediction  Variance 

J.  Andrew  Royle*  and  Doug  Nychka 


Abstract-  Stationarity  of  a  random  field  is  an  important  assumption 
required  for  many  spatial  statistical  analyses.  A  part  of  this  assump- 
tion is  homogeneity  (translation  invariance)  of  the  correlation  function. 
False  assumption  of  homogeneity  leads  directly  to  misspecification  of 
the  spatial  covariance  and  hence  the  usual  kriging  predictor  is  no  longer 
BLUE.  In  this  paper  we  examine  the  effect  of  homogeneous  misspec- 
ification of  a  heterogeneous  correlation  function  on  prediction  standard 
errors  using  two  cases  from  a  class  of  heterogeneous  models  that  are 
spatially  weighted  mixtures  of  homogeneous  random  fields.  It  is  found 
that  misspecification  leads  to  an  average  loss  of  efficiency  in  predic- 
tion by  up  to  20%  for  extreme  cases,  and  that  misspecification  of  the 
prediction  variance  generally  leads  to  conservative  prediction  intervals. 


1.  INTRODUCTION 

Kriging  is  a  commonly  used  method  for  making  spatial  predictions.  One  of  the 
primary  reasons  for  the  use  of  kriging  over  other  methods  of  spatial  prediction 
is  that  prediction  standard  errors  are  easily  obtained  for  kriging  and  that  the 
kriging  predictor  is  BLUE.  Unfortunately,  the  BLUE  property  holds  only  under 
the  assumption  that  the  spatial  covariance  is  known,  which  is  not  likely  to  be 
the  case  in  practice  (i.e.  there  is  generally  a  misspecification  of  the  covariance 
function). 

Kriging  predictions  are  known  to  be  fairly  robust  to  misspecification  of  the 
spatial  covariance  structure  (see  Cressie  and  Zimmerman,  1992).  Several  authors 
have  addressed  the  problem  of  stability  of  the  kriging  variance.  In  particular,  Stein 
(1988,  1989)  examined  the  effect  of  misspecification  of  the  covariance  structure 
on  prediction  variance.  For  a  particular  (fairly  general)  model,  he  derived  a  bound 

"Both  Authors:  Department  of  Statistics,  North  Carolina  State  University,  Raleigh  NC 


137 


on  the  ratio  of  the  optimal  prediction  variance  to  the  prediction  variance  for  the 
predictor  based  on  a  misspecified  covariance.  More  generally,  he  has  shown  that 
the  misspecified  prediction  variance  is  asymptotically  equivalent  to  the  minimum 
prediction  variance,  assuming  the  true  and  misspecified  covariance  functions  are 
compatible  (see  Stein  and  Handcock  (1989)  for  a  discussion  of  this  condition). 

Here  we  are  interested  in  the  departure  of  a  random  field  from  the  assumption 
of  translation  invariance  of  the  spatial  correlation.  Failure  to  detect  departure  from 
this  assumption  leads  directly  to  misspecification  of  the  covariance  function,  and 
we  quantify  the  effects  of  this  misspecification  on  prediction  variance  under  a  class 
of  heterogeneous  random  field  models. 

2.  STATIONARITY  AND  HETEROGENEITY 

Let  Z  =  {Z(x)  :  x  e  $d}  be  a  random  field  in  d-dimensions,  from  which  we 
observe  z  =  {z(x{)  :  1  <  i  <  p]  where  x  represents  a  vector  of  coordinate 
locations  in  space  (e.g.  x  =  (x\,X2)  in  -ft2).  Z  is  second-order  stationary  if 

E(Z(x))  =  fi  Vxeafc2 

Var(Z(x))  =  a2     Vz  £  ft2 

and 

Corr(Z{x),  Z(x'))  =  p{x  -  x')      Vx,  x'  €  $td 

Thus  under  second-order  stationarity  we  have  that  the  mean  and  covariance  do  not 
depend  on  the  location  of  the  points  in  space  (i.e.  translation  invariance  of  the 
first  and  second  moments  of  Z).  Concern  over  the  validity  of  the  stationarity 
assumption  generally  focuses  on  constancy  of  the  mean  and  variance,  however 
homogeneity  (translation  invariance)  of  the  spatial  correlation  is  also  implied  by 
stationarity.  A  heterogeneous  random  field  then,  is  one  in  which  the  correlation 
function  is  not  invariant  to  translation. 

2.1  A  Class  of  Heterogeneous  Random  Field  Models 

To  examine  the  effect  of  heterogeneity  on  prediction  variance  it  is  necessary  to 
specify  a  random  field  model  with  heterogeneous  covariance  structure.  Here  we 
consider  special  cases  from  the  class  of  models  that  are  spatially  weighted  mixtures 
of  homogeneous  random  fields.  The  2  component  mixture  is  given  by 

Z(x)  =  a(x)U{x)  +  b(x)V(x).  (1) 

Here  the  sampled  field  Z,  is  a  spatially  weighted  sum  of  the  homogeneous  random 
fields  U  and  V  with  covariance  functions  cu  and  cv  and  cross-covariance  cuv.  The 


138 


nature  of  the  spatial  weights  a(x)  and  b(x)  determines  the  manner  in  which  the 
covariance  between  two  points  depends  on  their  relative  locations.  Note  that  the 
covariance  between  Z  at  x  and  x'  under  (1)  is 

Cov(Z(x),Z(x'))    =   a(x)a(x')cu(x-x')  +  b(x)b(x')cv(x-x')  (2) 

+  (a(x)b(x')  +  a(x')b(x))cuv(x  -  x) 

It  is  obvious  that  the  covariance  between  two  points  depends  explicitly  on  their 
locations  x  and  x'  via  the  weights  a{x)  and  b(x),  and  hence  Z  is  a  heterogeneous 
random  field. 

Simple  cases  of  the  mixture  model  are  the  heterogeneous  variance  model: 
Z(x)  =  a(x)U(x),  where  Var(Z(x))  is  proportional  to  a(x)  and  the  hetero- 
geneous measurement  error  model:  Z(x)  =  U(x)  +  b(x)V(x)  where  V(x)  is 
measurement  error  with  constant  variance  and  Z[x)  has  measurement  error  (i.e. 
"nugget")  variance  proportional  to  b(x). 

When  the  spatial  weights  are  disjoint  indicator  functions  (i.e.  either  U(x)  or 
V(x)  are  observed  at  each  x)  and  cuv  =  0,  then  Z  is  homogeneous  within  subre- 
gions  and  Z  has  been  called  pseudo-stationary,  or  locally  stationary.  This  model  is 
equivalent  to  that  implied  by  partitioning  a  region  up  into  disjoint  (homogeneous) 
subregions,  a  common  solution  to  heterogeneity  in  classical  geostatistics  (Guttorp 
and  Sampson,  1994).  In  this  study,  we  consider  the  2  subregion  case  further 
(Model  7,  described  below).  A  similar  model  is  the  relative  variogram  model  of 
Cressie  (1991)  except  that  the  later  assumes  a  proportional  relationship  between 
the  covariance  and  mean  within  each  subregion. 

Taking  the  spatial  weights  to  be  overlapping  (but  not  constant)  produces  a 
reasonable  range  of  heterogeneous  behavior,  and  this  is  the  second  situation  that 
will  be  examined  here  (Model  2,  described  below).  Moving-window  kriging 
of  Haas  (1990)  implies  an  underlying  mixture  model  of  this  sort  but  with  >  2 
components.  Also  related  is  the  common  (homogeneous)  geostatistical  "linear 
model  of  coregionalization"  (Isaaks  and  Srivastava,  1989)  which  takes  constant 
weight  on  each  component  of  the  mixture  and  hence  the  covariance  is  just  a 
(weighted)  sum  of  2  or  more  covariance  functions. 

Let  Z  be  defined  over  the  region,  denoted  as  X,  containing  the  6  x  5  lattice  of 
sample  points  shown  in  Figure  1(a).  Define  subregions  of  X  as  S  —  {x  :  x2  <  3.5} 
and  Af  =  {x  :  x2  >  3.5}  and  hence  X  =  S  U  Af.  The  two  subregions  of  X  are 
delineated  by  the  broken  line  of  Figure  1(a).  The  2  heterogeneous  models  that  we 
consider  are  defined  as  follows: 

Model  1: 

Take  U  and  V  to  be  independent  (i.e.  cuv  =  0)  and  to  have  isotropic  expo- 
nential covariance  functions,  but  with  (possibly)  different  exponential  parame- 
ters. So,  cu(\\x  -  x'H)  =  e-l|x-r and  cv(\\x  -  x'\\)  =  Note  that 


139 


0U  =  —  log(pu(\))  and  9V  =  —log(pv{\))  so  that  we  may  describe  the  covariance 
functions  of  U  and  V  in  terms  of  the  "nearest-neighbor"  correlations  pu(l)  and 
pv(l).  These  are  the  correlations  between  two  points  separated  by  distance  1  in 
Figure  1(a).  The  weighting  scheme  is  defined  as 


and  b{x)  =  1  —  a(x).  This  case  is  relatively  uninteresting,  but  serves  as  a  "worst 
case"  scenario,  under  which  we  might  expect  misspecification  of  the  covariance 
structure  to  have  the  largest  effect.  The  heterogeneous  correlogram  of  Z  on  the 
6x5  lattice  (which  is  equivalent  to  the  covariogram  in  this  case)  is  shown  as 
the  solid  lines  in  Figure  1(b).  One  could  generalize  this  situation  to  allow  cross- 
correlation  between  U  and  V,  and  an  example  of  such  a  covariance  function  is 
shown  as  the  solid  lines  in  Figure  1(c).  Alternatively,  we  may  induce  a  similar 
heterogeneous  structure  by  overlapping  the  spatial  weights,  which  gives  us  our 
second  case. 


Taking  U  and  V  as  in  Model  7,  build  dependence  among  all  Z(x)  by  superimposing 
the  V  field  onto  the  U  field  (but  only  over  part  of  the  region,  producing  the 
heterogeneity).  Here  the  weighting  scheme  takes  a(x)  =1;      Vz  G  X  and 


Thus,  the  covariance  among  points  that  both  have  X2  €  S  contains  the  additive 
term  cv(\\x  —  x'\\).  The  covariance  among  all  other  points  is  simply  given  by 
cu(\\x  —  x'\\).  The  correlogram  of  Z  for  this  case  is  shown  as  the  solid  line 
in  Figure  1(d).  The  heterogeneous  models  shown  in  Figures  1(c)  and  (d)  are 
functionally  very  similar,  as  is  evident.  That  is,  they  both  build  dependence 
among  all  Z(x),  but  do  this  in  a  slightly  different  manner. 

We  will  define  the  misspecified  covariance  function,  cm,  to  be  the  least-squares 
projection  of  the  true  covariance  structure  to  a  homogeneous  exponential  covari- 
ance (i.e.  the  least- squares  fit  to  the  model  cm(||#  —  xf\\)  —  /^e-"*-37'"^2).  For 
Model  I  and  Model  2,  the  misspecified  homogeneous  covariance  functions  are 
shown  as  the  broken  lines  in  Figure  1(b)  and  (d),  respectively. 


Let  Z  be  mixture  of  two  homogeneous  Gaussian  random  fields  with  E(U(x))  = 
E(V(x))  =  0.  For  this  study,  we  choose  a  mixture  of  2  homogeneous  fields  for 
convenience  alone.  The  covariance  between  Z(x)  and  Z(x')  is 


Model  2: 


3.  STUDY  DESIGN 


Cov(Z(x),  Z(x'))  =  c(x,x'). 


140 


(a) 


(b) 


CO 

m 


9 


co 
d 


d 


o 
d 


2  4  6  8 


x1 


lag  distance 


(c) 


(d) 


00 

d 


d 


o 
d 


00 

d 


d 


o 
d 


lag  distance 


lag  distance 


Figure  1 :  (a)  -  Location  of  30  sample  points  used  in  studying  the  effect  of 
heterogeneous  correlation.  X  is  the  region  containing  these  points  and  the 
subregions  S  and  N  are  delineated  by  the  broken  line,  (b)  -  Heterogeneous 
correlation  of  30  points  under  Model  1  with  cuv  =  0.  (c)  -  Heterogeneous 
correlation  of  30  points  under  Model  I  with  puv{  1 )  =  .55.  (d)  -  Heterogeneous 
correlation  of  30  points  under  Model  2.  Broken  line  in  all  panels  is  the  mis- 
specified  homogeneous  correlation  function.  For  all  cases,  /)u(l)  =  .8  and 
Ml)  =  -3 


141 


For  Z(x)  corresponding  to  the  p  locations,  x,  at  which  the  random  field  is  measured 
we  have  the  p  x  p  matrix  Var(Z(x))  =  C  where  CtJ  =  c(x{,Xj).  For  any 
xQ  €  X,  we  have  the  ^-vector  Cov(Z(x0),  Z(x))  =  c0.  The  p  locations  where  Z 
is  observed  are  given  by  the  6  x  5  lattice  shown  in  Figure  1(a). 
For  x0  e  A\  the  kriging  predictor  of  Z(x0)  is 

Z(x0)  =  A'z 

where  the  vector  A  are  the  "kriging  weights",  and  are  computed  using  the  true 
covariance  function  defined  by  Model  1  or  Model  2  or  the  misspecified  covariance 
structure  determined  by  cm.   These  are  the  optimal  or  misspecified  predictor, 
respectively,  with  kriging  weights  A'  and  X'm . 
The  prediction  variance  is  given  by 

Var(Z(x0)  -  Z(x0))  =  1  -  2A'c0  +  A'CA 

and  the  prediction  standard  error  (PSE)  is  the  square  root  of  the  prediction  variance. 
The  prediction  variance  for  either  the  optimal  or  the  misspecified  predictor  may  be 
based  on  either  the  true  heterogeneous  covariance  function,  c,  or  the  misspecified 
homogeneous  covariance  function,  cm.  For  the  PSE  associated  with  the  prediction 
of  Z{x0),  3  different  estimates  will  be  denoted  as  PSEopt(x0),  PSEr(xQ),  and 
MPSE(x0).  The  3  estimators  are  defined  as: 

1.  PSEopt  -  The  optimal  estimator,  using  the  true  covariance  function  c  in  the 

predictor  and  prediction  variance  estimator. 

PSE2opt(x)  =  1  -  2A'c0  +  A'CA. 

2.  PSEr  -  The  correct  estimator  of  PSE  for  the  predictor  based  on  the  misspecified 

covariance,  cm. 

PSE2r{x)  =  \-2\'rnc0  +  \'rnC\rn. 

3.  MPSE  -  The  incorrect  estimator  of  PSE  for  the  predictor  based  on  the  mis- 

specified covariance,  cm . 

MPSE\x)  =  1  -  2A^c0,m  +  A'mCmAm. 

Note  that  the  kriging  weights,  and  covariance  are  indexed  by  m  for  the  misspecified 
covariance  function. 

The  estimator  PSEr  (the  subscript  r  for  reference)  serves  as  the  "best  possible" 
estimator  given  the  incorrect  predictor  (i.e.  using  the  predictor  which  is  not 
optimal,  but  using  the  correct  variance  for  the  prediction  error).  Since  PSEopt 


142 


is  the  optimal  estimator  of  PSE,  we  know  that  PSEr(x)  >  PSEopt(x),\/x  6  X. 
And,  since  MPSE  is  not  the  correct  PSE  estimator  for  the  predictor  based  on  the 
misspecified  covariance,  it  may  be  that  MPSE(x)  <  PSEopt(x). 

We  use  these  3  estimators  of  PSE  to  compute  prediction  standard  errors  on  a 
40  x  40  grid  of  points  over  the  region  containing  the  30  sample  points  of  Figure  1  (a). 
The  prediction  standard  errors  are  computed  using  only  the  30  sample  points,  and 
we  examine  interesting  summaries  of  the  3  PSE  estimators.  In  particular,  define 
the  average  efficiency  of  PSEr  to  PSEopt  as 


This  is  simply  a  measure  of  the  relative  precision  of  the  misspecified  predictor  to 
the  optimal  predictor.  Similarly,  to  quantify  the  worst  precision  over  X  define  the 
efficiency  of  the  worst  prediction  as 


These  summaries  provide  an  assessment  of  the  predictor  based  on  the  misspeci- 
fied covariance  function.  That  is,  under  misspecification  and  assuming  the  correct 
prediction  variance  is  used,  they  provide  insight  into  the  variability  of  the  predic- 
tions relative  to  the  optimal.  However,  supposing  the  heterogeneous  covariance 
structure  is  misspecified,  then  the  likely  estimator  of  prediction  standard  error  is 
MPSE  and  a  useful  measure  of  the  cost  of  misspecification  is  the  actual  coverage 
probability  for  a  95%  confidence  interval  for  a  prediction  using  MPSE,  averaged 
over  all  predictions.  For  any  x  6  X,  the  ratio  M P S E(x) / P S Er(x)  is  the  relative 
inflation  in  the  width  of  a  confidence  interval  as  a  result  of  using  the  misspecified 
PSE  estimator.  The  coverage  probability  of  a  95%  confidence  interval  for  the 
prediction  is  CP{x)  =  Pr(Z{x)  <  MPSE(x)/PSEr{x)  x  1.96)  -  Pr(Z(x)  < 
-MPSE(x)/PSEr{x)  x  1.96).  Hence,  if  MPSE(x)/PSEr{x)  =  1,  the  cov- 
erage probability  is  0.95.  The  average  coverage  probability  over  X  is 


Values  of  ACP  greater  than  0.95  indicate  conservative  confidence  intervals  (i.e. 
less  precise  predictions)  and  for  ACP  less  than  0.95,  predictions  are  too  precise. 

These  summaries  are  computed  for  values  of  pu(  1 )  =  .  1 ,  .2, . . . ,  .9  and  pv(  1 )  = 
.  1 ,  .2, . . . ,  .9,  for  each  of  the  2  models  defined  in  Section  2. 1  Recall  that  pu(  1 )  and 
pv(\)  are  the  nearest-neighbor  correlations  of  the  homogeneous  U  and  V  fields 
which  comprise  the  heterogeneous  mixture  model.  Thus,  for  each  heterogeneous 


Eff(r,opt)  = 


^5xPSE2opt{x)dx 


143 


case,  there  are  81  situations,  ranging  from  very  different  dependence  structure 
(e.g.  =  .1  and^(l)  =  .9),  to  identical  (but  independent)  structure  (pu(l)  = 
Pv(l),PuvW  —  0)-  Efficiency  and  ACP  will  be  presented  as  contour  plots  as  a 
function  of  pu(S)  and  pv(l). 

4.  RESULTS  AND  CONCLUSIONS 

Figure  2(a)  indicates  that  the  efficiency  of  prediction  is  effected  very  little  by 
misspecification  of  the  heterogeneous  covariance  structure  under  Model  1.  That 
is,  the  mean  prediction  efficiency  is  near  1  except  for  high  spatial  dependence 
(nearest-neighbor  correlations  >  0.8)  in  which  case  the  prediction  standard  errors 
are  only  20%  larger  for  predictions  based  on  the  misspecified  predictor.  For  the 
worst  prediction  under  misspecification  (Effmax),  Figure  2(e)  indicates  inflation 
of  PSE  by  20  to  120%  for  levels  of  spatial  correlation  >  0.50.  Since  it  is  unlikely 
that  one  would  use  PSEr  as  the  estimator  for  prediction  standard  error  if  the 
heterogeneous  covariance  were  misspecified,  perhaps  the  more  meaningful  result 
is  that  concerning  the  average  coverage  probability  using  MPSE.  The  ACP  is 
given  in  Figure  2(b),  where  it  is  seen  that,  using  MPSE  to  estimate  PSE,  one 
achieves  very  conservative  confidence  intervals  for  most  situations  under  Model  1. 
That  is,  the  ACP  for  a  95%  confidence  interval  is  generally  greater  than  0.95,  even 
approaching  1  for  high  spatial  dependence. 

For  Model  2,  Figure  2(c)  indicates  efficiency  of  the  misspecified  predictor 
near  1  and  hence  the  misspecified  spatial  predictor  is  nearly  optimal.  In  contrast 
to  Model  1,  Figure  2(f)  indicates  that  for  the  worst  predictions  made  under  mis- 
specification of  Model  2  the  increase  in  maximum  PSE  for  this  case  is  only  20 
to  40%  for  very  high  levels  of  spatial  correlation  (nearest-neighbor  correlations 

>  .80)  (compare  with  Figure  2(e)).  Also  in  contrast  to  Model  7,  the  ACP  using  the 
misspecified  estimator  of  PSE  (MPSE)  under  Model  2  (Figure  2(d))  is  generally 
less  than  0.95  (i.e.  liberal  coverage).  This  occurs  when  nearest-neighbor  correla- 
tions of  the  overlapping  U  and  V  are  relatively  very  different  (e.g.  /o«(l)  =  .6  and 
pv(\)  =  .2).  For  high  and/or  similar  nearest-neighbor  correlations,  the  ACP  is 

>  0.95  indicating  conservative  coverage  of  the  confidence  intervals  for  prediction. 

By  concentrating  on  a  particular  class  of  heterogeneous  models  for  spatial 
correlation,  defined  to  be  a  spatially  weighted  mixture  of  homogeneous  random 
fields  with  exponential  covariance  functions,  we  have  been  able  to  quantify  the 
effect  of  departure  from  homogeneity  on  spatial  prediction  uncertainty.  One 
problem  with  studying  the  effect  of  heterogeneity,  is  that  it  is  not  clear  how  to 
describe  the  phenomenon  in  a  manner  so  as  to  allow  generalization  of  results, 
such  as  those  presented  here,  to  arbitrary  covariance  structure.  Our  choice  of  the 
mixture  model  is  fortuitous  since  this  is  a  heterogeneous  analog  to  the  class  of 
models  examined  by  Stein  (1989)  which  are  of  the  form 

Cov(Z(s),  Z(s'))  =  a\cu(s  —  sf)  4-  b\cv(s  —  s) 


144 


(a):  Model  1  Eff(r.opt) 


(b):  Model  1  ACP 


(c):  Model  2  Eff(r.opt) 


(d):  Model  2  ACP 


0  2  0  «  0  6  OS 


(e):  Model  1  Effmax(r.opt) 


(f):  Model  2  EFFmax(r.opt) 


Figure  2:  (a)  Mean  efficiency  of  the  misspecified  predictor  under  Model  1. 
(b)  ACP  under  Model  1  for  95%  confidence  intervals  using  the  estimator  of 
prediction  standard  error  (MPSE)  based  on  the  misspecified  (homogeneous) 
covariance.  (c)  Mean  efficiency  of  the  misspecified  predictor  under  Model  2. 
(d)  ACP  using  MSPE  under  Model  2.  (e)  Efficiency  of  the  least-precise  pre- 
diction for  the  misspecified  predictor  under  Model  1.  (f)  Efficiency  of  the 
least-precise  prediction  under  Model  2.  Axes  are  nearest-neighbor  correla- 
tions of  the  U  and  V  fields. 


145 


and  were  used  by  him  to  study  the  effect  of  misspecification  (of  stationary)  covari- 
ance  structure,  and  thus  it  is  possible  that  results  such  as  this  may  be  extended  to 
include  heterogeneity  of  the  form  described  here,  but  with  an  arbitrary  covariance 
model  (i.e.  not  restricted  to  the  exponential  class).  This  is  an  area  of  future  work. 


REFERENCES 


Cressie,  N.A.C.  (1991).  Statistics  for  Spatial  Data.  Wiley,  New  York,  NY. 

Cressie,  N.A.C.  and  D.L.  Zimmerman  (1992).  On  the  Stability  of  the  geostatistical 
method.  Math.  Geol.  24(1),  45-60. 

Guttorp,  P.  and  P.D.  Sampson  (1994).  Methods  for  estimating  heterogeneous 
spatial  covariance  functions  with  environmental  applications,  in  G.P.  Patil  and 
C.R.  Rao,  eds.,  Handbook  of  Statistics  Volume  12,  661-689 

Haas,  T.C.  (1990).  Lognormal  and  moving  window  methods  of  estimating  acid 
deposition.  J.  Amer.  Stat.  Assoc.  85,  950-963. 

Isaaks,  E.H.  and  R.M.  Srivastava  ( 1 989).  An  Introduction  to  Applied  Geostatistics. 
Oxford  University  Press. 

Stein,  M.L.  (1988).  Asymptotically  efficient  prediction  of  a  random  field  with 
misspecified  covariance  function.  Ann.  Stat.  16(1),  55-63. 

Stein,  M.L.  (1989).  The  loss  of  efficiency  in  kriging  prediction  caused  by  mis- 
specifications  of  the  covariance  structure,  in  M.  Armstrong,  ed.,  Geostatistics. 
Kluwer  Academic  Publishers,  273-282. 

Stein,  M.L.  and  M.S.  Handcock  (1989).  Some  asymptotic  properties  of  kriging 
when  the  covariance  function  is  misspecified.  Math.  Geol.  21,  171-190. 


146 


Further  Explorations  of  Relationships 
between  Semi-Variogram  and  Spatial 
Autoregressive  Models 


Daniel  A.  Griffith,  Larry  J.  Layne,  and  Philip  G.  Doyle2 


Abstract. — This  paper  continues  the  forging  of  a  common  foundation  for  geo-  and 
spatial  statistics.  To  date  partial  conceptual  correspondence  has  been  established 
between  the  conditional  autoregressive  (CAR)  and  exponential  semi-variogram 
models,  the  simultaneous  autoregressive  (SAR)  and  Bessel  function  semi-variogram 
models,  and  the  moving  average  (MA)  and  linear  semi-variogram  models,  when 
directional  bias  is  not  present.  The  exploratory  numerical  work  summarized  in  this 
paper  addresses  the  issue  of  whether  or  not  these  articulations  are  preserved  when 
directional  bias  is  present  in  the  latent  spatial  autocorrelation,  given  symmetric  spatial 
dependency.  In  doing  so,  impacts  of  edge  effects  are  examined. 


INTRODUCTION 

Generally  speaking,  spatial  statistics  is  concerned  with  the  statistical  analysis  of  geo- 
referenced  (or  locationally  tagged)  data,  and  seeks  to  exploit  the  redundant  information 
represented  by  spatial  autocorrelation  latent  in  such  data.  Two  principal  thrusts  of  spatial 
analysis  have  been  (1)  interpolation  supporting  visualization  of  surfaces,  and  (2)  data 
analysis  supporting  statistical  description  and  inference.  One  of  the  seminal  visualization 
efforts  was  the  famous  SYMAP  computer  program,  which  sought  to  furnish  low-resolu- 
tion, coarse  visualization  of  geographic  surface  representations  (contour  maps  or  perspec- 
tive surfaces)  based  upon  measurements  made  at  a  relatively  small  number  of  dispersed 
locations.  Paralleling  the  development  and  dissemination  of  SYMAP  was  a  body  of  work 
that  focused  on  exploiting  more  localized  spatial  dependence  latent  in  geo-referenced  data, 
which  has  become  known  as  the  theory  of  regionalized  variables,  or  geo-statistics. 
Matheron  initiated  this  pioneering  work  in  the  1960s,  and  Cressie  (1991)  recently  pub- 
lished a  state-of-the-art  overview  of  geo-statistics.  Meanwhile,  spatial  statistics  research, 
especially  characterized  by  Cliff  and  Ord's  (e.g.,  1981)  work,  focusedattention  on  the  need 
to  employ  spatially  adjusted  linear  statistical  models.  Thus,  while  both  geo-  and  spatial 


1This  research  was  supported  by  the  National  Science  Foundation,  research  grant  #  SBR-9507855. 

Professor,  Research  Associate,  and  undergraduate  research  assistant  (through  an  NSF  REU  supplement 
grant),  respectively,  Department  of  Geography,  Syracuse  University,  144  Eggers  Hall,  Syracuse, 
NY  13244-1090. 


147 


statistics  support  the  visualization  of  geo-referenceddata,  spatial  statistics  goes  beyond  this 
task  by  enabling  the  drawing  of  sound  statistical  inferences  from  geo-referenced  data. 

While  Davis  and  McCullagh  (1975)  began  forging  a  common  foundation  for  these  two  sub- 
disciplines,  such  a  connection  has  continued  to  remain  elusive.  The  prhcipal  objective  of 
the  research  findings  reported  in  this  paper  is  to  pursue  this  end.  Its  practical  importance 
partly  arises  within  the  U.S.  E.P.A.  Environmental  Monitoring  and  Assessment  Program 
(EMAP)  project,  a  major  national  endeavor  in  which  statistics  is  being  appliedto  environ- 
mental issues  as  well  as  monitoring  and  analysis  undertakings,  and  partly  in  remediation 
work  dealing  with  environmental  contaminants.  Concepts  from  both  geo-  and  spatial 
statistics  are  necessary  for  successful  engagement  in  these  two  scientific  pursuits.  Articu- 
lation of  a  common  foundation  for  geo-  and  spatial  statistics  will  better  enable  consistent 
statistical  application  and  analysis  of  the  massive  amounts  of  geo-referenced  environmental 
data  becoming  available.  By  doing  so,  geographic  visualization  and  analysis  will  be  able 
to  more  closely  go  hand  in  hand,  and  inteipolated  maps  will  be  better  able  to  be  analyzed 
utilizing  consistent  spatial  statistical  model  specifications. 

Communalities  Between  Geo-  and  Spatial  Statistics 

Historically,  traditional  statistical  analyses  have  been  concerned  with  datareduction  and 
summarization.  In  contrast,  time  series  analysis  frequently  leads  to  forecasting,  whereas 
spatial  series  analysis  often  leads  to  interpolation,  two  data  explosion  operations.  Geo-  and 
spatial  statistics  similarly  fixate  on  this  interpolation  issue  through  the  "missing  data" 
problem.  Martin  (1984)  has  outlined  the  exact  maximum  likelihood  estimates  of  missing 
values  for  incomplete  geographic  data  sets  from  a  spatial  statistics,  autoregressive  model- 
ling perspective.  The  resulting  equations,  as  demonstrated  in  Griffith  (1993),  are  exactly 
those  best  linear  unbiased  estimates  obtained  with  kriging  (see  Christensen,  1991),  but  in 
terms  of  the  inverse  covariance  matrix  rather  than  the  covariance  matrix  itself.  This  link 
initially  provided  a  conspicuous  piece  of  evidence  suggesting  close  relationships  between 
semi-variogram  and  spatial  autoregressive  models,  and  furnished  the  motivation  for 
Griffith  and  Csillag  (1993)  to  numerically  explore  this  relationship. 

Algebraically  speaking,  then,  estimation  of  missing  data  in  spatial  statistics  (in  order  to 
avoid  maps  with  holes)  and  kriging  as  an  interpolator/predictor  are  exactly  equivalent, 
which  connects  kriging  to  the  E-M  algorithm  literature  of  statistics.  Both  numerical  and 
graphical  evidence  exist  to  indicate  that  the  spatial  statistical  conditional  autoregressive 
(CAR)  and  geo-statistical  exponential  semi-variogram  models  are  linked.  Graphical 
evidence  exists  that  couples  the  spatial  statistical  moving  average  (MA)  and  geo-statistical 
linear  semi-variogram  models.  And,  both  numerical  evidence  and  theoretical  arguments 
signal  a  relationship  between  the  spatial  statistical  simultaneous  autoregressive  (SAR)  and 
the  geo-statistical  Bessel  function  (first  order,  second  kind)  semi-variogram  models.  The 
spectral  density  function  of  the  SAR  model  may  be  written  as: 

Pki  =  {o/*  o/n  cos(kBi)cos(iej/[i  -  2PpCas(0O  - 
2Pqcas(02)]2  do,  d02}/{o/-  j«  [i  -  2Ppco5(01)  - 2Pqcas(02)]-2  de,  d©2}.  (i) 


148 


Christensen  (1991)  summarizes  the  status  of  this  model  by  noting  that  Whittle  (1954)  has 
shown  that  the  Bessel  function  can  arise  naturally  in  two-dimensional  covariations, and  the 
exponential  semi-variogram  model  is  a  special  case  of  it  (Ripley  1981).  The  relevant 
functional  form,  where  K,  is  a  modified  Bessel  function,  is  given  by: 

1  -  pkJ  =  c0  +  c, [  1  -  (distance/range)xK,(distance/range)]  .  (2) 

When  considering  the  appropriateness  of  this  latter  function  as  a  descriptor  of  geographic 
covariation,  numerical  evidence  (computed  with  Heuvelink's  weighted  least  squares 
software)  based  on  an  isotropic  dependency  structure  indicates  that  the  two  critical  features 
leading  to  a  Bessel  function  semi-variogram  model  specification  appear  to  be  intensity  and 
degree  of  articulation  of  a  geographic  structure  (see  Table  1). 


Table  1.  Bessel  function  semi-variogram  fit  of  auto  regressive  models  for  geographic  structures. 


Geographic 
articulation 

Autoregressive 
conditional 

model 

simultaneous 

c0 

c, 

range 

SSD/SST 

c0 

C] 

range 

SSD/SST 

square 

0.350 

0.617 

2.56 

0.007 

0.005 

0.994 

8.27 

0.000 

hexagon 

0.443 

0.489 

2.19 

0.041 

queen 

0.442 

0.520 

2.88 

0.004 

0.011 

0.988 

10.1 

0.000 

Of  note  is  that  the  exponential  semi-variogram  model  (which  actually  is  a  Bessel  function 
of  order  Vi)  furnishes  a  better  description  of  the  CAR  cases  than  does  the  particular 
modified  Bessel  function  estimated  here.  An  SAR  structure,  which  involves  second-order 
dependencies,  is  better  characterized  by  the  popular  Bessel  function  than  is  the  CAR 
structure,  which  involves  only  first-order  dependencies.  Furthermore,  relying  upon  an 
analogy  with  chess,  the  spatial  linkage  structure  for  the  queen's  set  of  neighbors  (those 
areal  units  sharing  any  common  boundary  point)  is  better  characterized  by  this  Bessel 
function  than  is  the  rook's  set  of  neighbors  (only  areal  units  sharing  a  common  boundary 
of  non-zero  length).  Moreover,  as  the  spatial  random  field  increases  in  scope  (i.e.,  moving 
from  a  rook's  to  a  queen's  move  dependency  field),  the  SAR  model  more  strongly  relates 
to  the  Bessel  function  semi-variogram  model.  That  is,  the  discrete  setting  more  closely 
resembles  the  continuous  setting.  This  tendency  holds  important  implications  f>r  the  U.S. 
E.P.A.  EMAP  database  currently  being  constructed,  which  is  being  geo-referenced  in 
terms  of  a  hexagonal  tessellation  (comprised  of  approximately  12,600  hexagons)  superim- 
posed upon  the  continental  U.S. 

Numerical  evidence  presented  in  Table  1  was  taken  from  exploratory  computer  experi- 
ments involving  the  following  three  general  steps: 

Step  l :  calculation  of  theoretical  spatial  correlations  at  various  lags,  for  a  regular  square 
tessellation,  based  upon  spatial  autoregressive  model  specifications 

Step  2:  fitting  of  selected  geo-statistical  semi-variogram  models  to  the  two-dimensional 
correlograms  obtained  in  Step  1  using  non-linear  least  squares 

Step  3:  graphical  superimposition  of  the  estimated  weighted  least  squares  curves  on  the 
correlograms  obtained  in  Step  1  using  Heuvelink's  computer  software  or  custom- 
ized SAS  code. 


149 


The  findings  are  restricted  to  translations  from  spatial  statistical  to  geo-statistical  model 
specifications,  spotlighting  the  three  overwhelmingly  most  popular  spatial  autoregressive 
models  (i.e.,  the  MA,  CAR  and  SAR  models). 

Fortifying  the  Autoregressive-geo-statistical  Linkages 

Research  findings  summarized  here  pertain  to  a  further  articulation  of  a  common  founda- 
tion for  geo-  and  spatial  statistics.  One  motivation  for  pursuing  this  work  is  the  need  for 
establishing  a  common  umbrella  for  a  growing  body  of  seemingly  disparate  work  (e.g., 
Rey,  Getis  and  Bortman,  1994;  Griffith,  1992;  Little  and  Rubin,  1987).  The  extension 
outlined  here  builds  on  results  reported  by  Griffith  (1993)  and  Griffith  andCsillag  (1993). 
Two  principal  themes  have  been  examined.  One  concerns  non-stationary,  multi-parameter 
spatial  autoregressive  models.  Martin  (1990)  notes  that  the  use  of  even  bi-parametric 
spatial  autoregressive  inverse-covariance  matrix  specifications  is  extremely  rare  in  applica- 
tions, although  well-established  for  very  special  cases.  This  avoidance  of  the  two-parame- 
ter model  in  part  is  due  to  numerical  difficulties  affiliated  with  its  estimation.  The  critical 
simplicity  feature  of  a  specification  is  that  the  pair  of  geographic  weights  matrices  either 
are  commutative  or  can  exploit  the  matrix  algebra  notion  of  similarity.  In  contrast,  specifi- 
cation and  estimation  of  directional  semi-variogram  functions  is  quite  common. 

The  second  theme  concerns  boundary,  or  edge,  effects.  Griffith  (1983)  addresses  the 
general  problem  of  edge  effects  in  quantitative  geographical  analyses,  underscoring  their 
seriousness.  Rathbun  (1994)  provides  an  excellent  example  of  the  type  of  distortion 
overlooked  edge  effects  can  create  and  showed  that  properly  handling  them  allowed  better 
interpolation  of  salinity  in  Chesapeake  Bay.  In  addition,  results  reported  in  Griffith  and 
Csillag  (1993)  exhibit  an  affinity  between  the  magnitude  of  prevailing  spatial  dependencies 
and  the  dominance  of  regional  edge  effects.  This  factor  emerges  because  semi-variogram 
models  directly  deal  with  the  covariance  matrix  for  a  set  of  observations,  whereas  spatial 
autoregressive  models  deal  with  the  inverse  of  this  matrix,  and  the  procedure  of  matrix 
inversion  is  what  amplifies  and  propagates  edge  effects.  To  date  these  effects  have  defied 
the  formulation  of  general  correction  techniques.  Insight  into  this  edge  effects  phenome- 
non is  gained  here  by  holding  the  degree  of  spatial  dependence  constant,  and  changingthe 
size  of  the  region  within  which  estimation  is  conducted  (see  Table  2). 

Table  2.  Bessel  function  parameter  estimates  for  queen's  case,  isotropic  SAR  specification, 


p  =  0.2491. 


Maximum  lag 

c0 

Ci 

range 

relative  SS 

5 

0.0093 

0.906 

9.35 

0.000 

10 

0.0103 

0.965 

9.86 

0.000 

15 

0.0109 

0.978 

10.00 

0.000 

25 

0.0116 

0.987 

10.10 

0.000 

35 

0.0119 

0.987 

10.10 

0.000 

5000 

0.0200 

0.987 

10.50 

0.000 

Theoretically  one  would  expect  c0  =  0  and  Cj  =  1 .  Notice  that  as  the  region  used  for 


150 


parameter  estimation  purposes  increases  in  size,  the  fit  of  the  Bessel  function  remains  near- 
perfect,  while  the  parameter  estimates  themselves  change.  The  range  of  the  spatial 
autocorrelation  field  increases  to  some  upper  limit  and  q  approaches  1.  Somewhat 
surprising  is  that  Cq  (the  nugget  effect)  does  not  move  toward  zero,  which  may  be  indica- 
tive of  a  liability  associated  with  using  a  non-linear  model  specification  and/or  weighted 
least  squares  estimation. 

Two-parameter  Autoregressive  Functions  and  Directional  Semi-variogram 

Models 

Theoretical  correlograms  can  be  constructed  for  the  bi-parametric  spatial  autoregressive 
model  matched  with  a  regular  square  tessellation  for  the  rook's  set  of  neighbors  spatial 
linkages  since  the  closed  form  spectral  density  function  is  known.  As  noted  previously, 
specification  and  estimation  of  directional  semi-variogram  functions  is  common.  The 
theoretical  two-dimensional  correlograms  generated  by  these  equations  do  not  need  to  be 
rotated,  as  their  major  and  minor  axes  coincide  with  the  corresponding  horizontal  and 
vertical  axes. 

Results  of  bi-parametric  semi-variogram  model  fits  are  reported  in  Table  3.  As  has  been 
found  for  the  isotropic  case,  the  exponential  semi-variogram  model  relates  almost  perfectly 
to  the  spatial  statistical  CAR  model  case,  and  the  Bessel  function  semi-variogram  model 
relates  almost  perfectly  to  the  spatial  statistical  SAR model  case.  For  all  cases  explored 
here,  the  relative  residual  sums  of  squares  essentially  equals  zero.  This  indicates  that  the 
linkages  already  established  between  the  geo-statistical  and  spatial  statistical  models 
persist  in  the  presence  of  anisotropy.  Differences  in  the  computed  directional  ranges  are 
consistent  with  differences  between  the  directional  autoregressive  parameters  pp  and  pq. 
Furthermore,  an  inspection  of  the  change  in  semi-variogram  parameter  estimates  as  the 
size  of  the  region  under  study  increases  reveals  the  presence  of  edge  effects,as  it  did  in  the 
isotropic  situation.  These  edge  effects  become  more  pronounced  as  the  two  levels  of 
spatial  autocorrelation  increase. 

Table  3.  Parameter  estimates  and  relative  sums  of  squares  (error  sums  of  squares/corrected  total 
sums  of  squares,  or  SSE/CTSS)  for  the  CAR  and  SAR  models  for  4  different  lattice  sizes. 

Autoregressive  model 


conditional 

simultaneous 

lattice 

p,=  0.05 

0.175 

p!  =  0.050 

p!=  0.05 

p^  0.175 

Pt=  0.050 

size 

parameter 

p,  =  0.25 

p,  =  0.320 

p7  =  0.445 

p,  =  0.25 

p,  =  0.320 

p,  =  0.445 

4x4 

0.00000 

0.13062 

0.05397 

0.02864 

0.00922 

0.00645 

c, 

1.00045 

0.79249 

0.89910 

0.97118 

0.93289 

0.92101 

b, 

0.35341 

1.52411 

0.95340 

0.35638 

3.84913 

1.93552 

b2 

0.76714 

2.13413 

3.08842 

0.84330 

5.25147 

6.14246 

SSE/CTSS 

0.00328 

0.00338 

0.00328 

0.00225 

0.00013 

0.00059 

6x6 

C0 

0.00000 

0.18844 

0.09685 

0.02906 

0.01169 

0.01215 

c, 

1.00020 

0.76710 

0.87825 

0.97090 

0.96549 

0.96173 

0.35265 

1.80853 

1.07463 

0.35662 

4.00823 

2.07320 

151 


b2 

0.76649 

2.51674 

3.43566 

0.84382 

5.44961 

6.42030 

SSE/CTSS 

0.00309 

0.00416 

0.00554 

0.00204 

0.00006 

0.00041 

8x8  C0 

0.00000 

0.22557 

0.12686 

0.02915 

0.01335 

0.01553 

c1 

1.00011 

0.74630 

0.85815 

0.97083 

0.97558 

0.97252 

b1 

0.35239 

2.00544 

1.15218 

0.35666 

4.07235 

2.12743 

b2 

0.76625 

2.78110 

3.66908 

0.84392 

5.52885 

6.51884 

SSE/CTSS 

0.00302 

0.00509 

0.00704 

0.00197 

0.00004 

0.00032 

10  x  10  C0 

0.00000 

0.25132 

0.14723 

0.02919 

0.01454 

0.01771 

c1 

1.00007 

0.72972 

0.84279 

0.97080 

0.97932 

0.97589 

b, 

0.35227 

2.14681 

1.20164 

0.35668 

4.10434 

2.15386 

b2 

0.76615 

2.97021 

3.82763 

0.84396 

5.56817 

6.56468 

SSE/CTSS 

0.00298 

0.00594 

0.00804 

0.00193 

0.00003 

0.00027 

Theoretically  the  parameter  estimates  for  C0  and  C,  theoretically  should  equal  0  and 
1,  respectively.  These  are  approximately  the  values  estimated  involving  low  levels  of 
spatial  autocorrelation,  for  both  the  exponential  and  Bessel  function  semi-variogram 
models.  A  tendency  for  Cj  to  move  toward  1  is  apparent  for  the  intermediate  and  high 
levels  of  spatial  autocorrelation  in  both  model  cases.  Likewise  a  movement  toward  0  is 
detectable  for  C0  in  the  exponential  case  but  not  in  the  Bessel  function  case. 

Empirical  Case  Studies:  The  Mercer-Hall  Data  and  Pediatric  Blood  Lead  Levels 

in  Syracuse 

Two  empirical  data  sets  are  investigated  here.  The  first  is  the  Mercer-Hall  agricultural 
wheat-yield  field  plot  data.  These  data  were  collected  for  500  plots  that  formed  a  regular 
square  25-by-20  tessellation.  Cressie  (1991,  p.  454)  summarizes  these  data  and  their 
geographic  landscape  features.  The  second  is  pediatric  blood  lead  levels  in  the  City  of 
Syracuse  measured  during  the  two-year  period  4/1/92-3/31/94.  This  data  set  is  com- 
prised of  7,158  measurements  (out  of  15,277  cases  for  the  County),  some  of  which  are 
replicates,  spread  across  3,841  locations.  Twelve  children  had  measures  of  zero,  and 
for  convenience  were  dropped  from  the  analysis.  Locational  tags  for  these  measures 
were  obtained  by  address  matching  with  ARC/INFO  using  TIGER  files  for  Onondaga 
County,  NY  (with  about  an  80%  success  rate). 

The  Mercer-Hall  Data 

These  data  appear  to  be  approximately  normally  distributed  [P(Wilk-Shapiro  statistic 
<  0.98041)  =  0.0858].  Besag  (1974)  fit  a  bi-parametric  constant  mean  CAR  model  to 
these  data,  using  a  normalizing  approximation  proposed  by  Whittle  (1954),  and  obtained 
the  parameter  estimates  pp  =  0.368  and  pq  =  0.107.  Exact  maximum  likelihood  estima- 
tion (using  IMSL  subroutine  UMINF)  yields  the  estimates  pp  =  0.36445  and  pq  = 
0. 1 1388,  while  using  an  extension  of  the  Jacobian  approximation  outlined  by  Griffith  and 
Sone  (1995)  implemented  in  SAS  yields  the  estimates  pp  =  0.37634  and  pq  =  0.12480. 
Both  of  these  latter  solutions  began  iterating  from  the  OLS  values  of  pp  =  0.34045  and 
pq  =  0.13746.  A  mapping  of  the  two-dimensional  correlogram  corroborated  the  need 


152 


for  an  anisotropic  spatial  statistical  model  specification,  one  that  is  symmetric  and  bi- 
parametric  but  not  in  need  of  an  axis  rotation.  Restricting  estimation  of  the  exponential 
semi-variogram  to  10  lags,  and  using  weighted  least  squares  and  a  model  specification 
version  following  Haining  (1990,  p.  285),  rendered  the  parameter  estimates 

c0  cB  range  #1       range  #2  SSD/SST 

0.350       0.617         3.37411         0.74487  0.36600 

The  semi-variogram  plot  for  these  data,  upon  which  the  exponential  model  predictions  are 
superimposed,  is  indicative  of  both  a  CAR  and  a  bi-parametric  model  specification. 

The  Syracuse  Pediatric  Blood  Lead  Level  Data 

A  log-transformed  version  of  these  data  is  closer  to  being  approximately  normally 
distributed  [P(Wilk-Shapiro  statistic  <  0.9895)  <  0.01]  than  are  the  raw  data  themselves, 
and  they  also  deviate  far  less  from  the  homogeneous  variance  case.  Because  of  the 
massive  number  of  pairs  of  directional  distances  that  can  be  computed  for  this  dataset  (i.e., 
14,753,281)  attention  has  been  restricted  here  to  distances  less  tian  or  equal  to  0.2  km.  A 
two-dimensional  correlogram  for  this  range  of  distances  indicates  the  need  for  a  bi-para- 
metric model  specification  involving  an  axis  rotation.  Results  of  such  an  exercise  are 
contained  in  Table  4.  The  rotation  angle,  0,  corroborates  the  need  for  an  axis  rotation  in 


Table  4.  Semi-variogram  estimation  for  Syracuse  blood  lead  levels. 


Model 

C0 

c, 

range  #1 

range  #2 

relative  SS 

6 

CAR 

0.20396 

0.58339 

0.00192 

0.01154 

0.37078 

0.04518  7i 

SAR 

0.20832 

0.57789 

0.00070 

0.04549 

0.36326 

0.03979  7i 

this  empirical  case.  The  range  estimates  corroborate  the  need  for  a  bi-parametric  model 
specification.  Also,  the  SAR  model  furnishes  a  marginally  better  semi-variogram  model 
fit. 

CONCLUSIONS 

Additional  numerical  and  empirical  evidence  is  presented  in  this  paper  supporting  the 
establishment  of  links  between  the  CAR  and  exponential  semi-variogram  models,  and  the 
SAR  and  Bessel  function  semi-variogram  models.  This  new  evidence  reveals  that  these 
links  persist  in  the  presence  of  anisotropy.  Moreover,  these  articulations  are  preserved 
when  directional  bias  is  present  in  latent  spatial  autocorrelation,  given  directionally 
symmetric  spatial  dependency. 

REFERENCES 

Besag,  J.  1974.  Spatial  interaction  and  the  statistical  analysis  of  lattte  systems.  Journal  of 

the  Royal  Statistical  Society,  36:  192-225. 
Christensen,  R.  1991.  Linear  Models  for  Multivariate,  Time  Series,  and  Spatial  Data.  NY: 

Springer-Verlag. 


153 


Cliff,  A.,  and  J.  Ord.  1981.  Spatial  Processes.  London:  Pion. 
Cressie,  N.  1991.  Statistics  for  Spatial  Data.  NY:  Wiley. 

Davis,  J.,  and  M.  McCullagh  (eds.).  1975.  Display  and  Analysis  of  Spatial  Data.  NY: 
Wiley. 

Griffith,  D.  1983.  The  boundary  value  problem  in  spatial  statistical  analysis.  Journal  of 

Regional  Science,  23:377-387. 
Griffith,  D.  1992.  Estimating  missing  values  in  spatial  urban  census  data.  The  Operational 

Geographer,  10  (2):  23-26. 
Griffith,  D.  1993.  Advanced  spatial  statistics  for  analyzing  and  visualizinggeo-referenced 

Data.  International  Journal  of  Geographical  Information  Systems,  7:  107-123. 
Griffith,  D.  and  F.  Csillag.  1993.  Exploring  relationships  between  semi-variogram  and 

spatial  autoregressive  models.  Papers  in  Regional  Science,  72:  283-296. 
Griffith,  D.,  and  A.  Sone.  1995  Trade-offs  associated  with  normalizing  constant  compita 

tional  simplifications  for  estimating  spatial  statistical  models.  Journal  of  Statistical 

Computation  and  Simulation,  51:  165-183. 
Haining,  R.  1990.  Spatial  Data  Analysis  in  the  Social  and  Environmental  Sciences. 

Cambridge:  Cambridge  University  Press. 
Heuvelink,  G.  1993.  Error  Propagation  in  Quantitative  Spatial  Modelling.  Utrecht,  The 

Netherlands:  Faculteit  Ruimtelijke  Wetenschappen  Universiteit  Utrecht. 
Little,  R.,  and  D.  Rubin.  1987.  Statistical  Analysis  with  Missing  Data.  NY:  Wiley. 
Martin,  R.  1984.  Exact  maximum  likelihood  for  incomplete  data  from  a  correlated 

Gaussian  process.  Communications  in  Statistics:  Theory  and  Methods,  13:  1275-1288. 
Martin,  R.  1990.  Spatial  statistical  processes  in  geographic  modelling/^?  Spatial  Statistics: 

Past,  Present,  and  Future,  edited  by  D.  Griffith,  pp.  109-127.  Ann  Arbor,  MI:  IMaGe. 
Rathbun,  S.  1994.  Spatial  modelling  in  irregularly  shaped  regions:  kriging  estuaries,  paper 

presented  at  the  annual  joint  statistical  meetings,  August  13-18,  Toronto,  Ontario, 

Canada. 

Rey,  S.,  A.  Getis,  and  A.  Batman.  1994.  Spatial  modeling  approaches  for  the  estimation 
of  suppressed  geo-referenced  data,  paper  presentedat  the  41st  North  American  meetings 
of  the  Regional  Science  Association  International,  Niagara  Falls,  Canada,  November  17- 
20. 

Ripley,  B.  1981.  Spatial  Statistics.  NY:  Wiley. 

Whittle,  P.  1954.  On  stationary  process  in  the  plane.  Biometrika,  41:  434-449. 

BIOGRAPHICAL  SKETCH 

Dr.  Daniel  A.  Griffith  is  currently  Chair,  Department  of  Geography  at  Syracuse  Univer- 
sity. He  holds  a  Ph.D.  in  Geography  from  the  University  of  Toronto  (1978),  an  M.S.  in 
Statistics  from  The  Pennsylvania  State  University  (1985),  an  M.A.  in  Geography  (1972) 
and  a  B.S.  in  Mathematics,  both  from  Indiana  University  of  Pennsylvania. 

Larry  J.  Layne  is  a  Ph.D.  candidate  at  the  College  of  Environmental  Sciences  and 
Forestry,  S.U.N.Y.  He  holds  an  M.S.  in  Wildlife  and  Fisheries  from  South  Dakota  State 
University  (1988)  and  a  B.S.  in  Wildlife  from  Humboldt  State  University  (1983). 

Philip  G.  Doyle  is  a  senior  in  the  Department  of  Geography  at  Syracuse  University  and 
a  G.I.S.  Intern  for  Niagara  Mohawk  Power  Corporation. 


154 


A  Gamma-function  Model  for 
D-variate  Spectra  and  Cross-spectra 
for  Large  Scale  Frequency  Domain 
Simulation  of  Stationary  Random 

Functions  in  Rn 

Leon  E.  Borgman1  and  John  W.  Kern2 


Abstract.— Spatially  correlated  random  fields  are  used  to  model  phenomena 
ranging  in  scale  from  description  of  the  distribution  of  fibers  in  a  sheet  of 
paper  to  simulation  of  the  forces  of  ocean  waves  along  a  beach.  To  make  use 
of  this  framework  with  computationally  efficient  frequency  domain  methods, 
characterization  of  auto-  and  cross-spectra  are  required.  Univariate  and  cross- 
spectral  models  based  on  the  gamma  probability  density  function  are  proposed 
for  d-variate  random  vectors  in  n-dimensional  space  with  geometric 
anisotropy.  The  corresponding  class  of  spatial  covariance  functions  is  also 
derived.  Modeling  the  second  order  structure  between  pairs  of  random 
variables  is  shown  to  simplify  the  usually  tedious  lag  domain  modeling  of 
cross-covanance  functions. 


INTRODUCTION 

Spatially  correlated  random  fields  have  been  used  to  model  physical 
phenomena  ranging  in  scale  from  description  of  the  distribution  of  fibers  in  a  sheet  of 
paper  (Ahuja  and  Schachter,  1983),  to  simulation  of  the  forces  of  ocean  waves  along 
a  beach  (Borgman  et  al.  1994).  Prediction  techniques  usually  known  as  kriging, 
(Krige  1951),  Matheron  (1971)  and  Journel  and  Huigbregts  (1978)  are  used  to 
produce  best  linear  unbiased  predictions  at  unsampled  locations  conditioned  on 
known  data  .  In  contrast,  conditional  simulation  methods  (Journel  and  Isaaks  1984), 
(Borgman  et  al.  1984,  1993),  are  used  to  produce  an  ensemble  of  realizations  of  the 
random  process  where  the  correlation  structure  and  natural  variability  of  the  original 
process  are  preserved  and  known  data  are  interpolated. 

Conditional  simulations  can  be  produced  through  space  domain  or  frequency 
domain  methods.  However,  for  large  scale  regional  simulations  (100  by  100  grids  or 
more),  the  space  domain  methods  are  slow  and  require  large  amounts  of  computer 
memory  to  execute.    To  overcome  the  limitations  of  space  domain  simulation 

1  Statistician.  Department  of  Statistics  and  Geology  Geophysics,  University  of  Wyoming.  I^aramie  WY. 
1  Statistician.  Western  EcoSystems  Technology  Inc.  Cheyenne  WY. 


155 


techniques,  the  faster  less  memory-intensive  frequency  domain  methods  were 
developed,  allowing  simulation  of  larger  spatially  correlated  random  fields  at  much 
greater  speeds. 

To  obtain  the  spectral  density  models  required  for  frequency  domain  techniques, 
covariance  functions  developed  for  use  in  space  domain  applications  such  as  the 
spherical  model,  hole  effect  model  or  members  of  the  Matern  class  (Matern  1 960), 
have  typically  been  Fourier  transformed  to  frequency  domain.  In  general,  these 
spectral  density  models  have  little  intuitive  connection  with  space  domain  processes 
and  in  some  cases  demonstrate  pathologies  in  practice. 

With  the  exception  of  Goulard  and  Voltz  (1992),  there  has  been  little  development 
of  cross  covariance/spectral  density  models  for  simulation  of  multivariate  correlated 
random  fields.  Since  few  environmental  or  ecological  problems  are  univariate  in 
nature,  this  lack  of  available  multivariate  models  has  limited  application  of  conditional 
simulation  techniques  to  these  areas. 

We  propose  a  class  of  univariate  and  multivariate  spectral/covariance  models 
suitable  for  simulation  of  regional  scale  processes.  The  class  of  models  is  flexible  and 
easily  interpreted  using  properties  of  the  individual  univariate  processes.. 

UNIVARIATE  SPECTRAL  MODEL  WITH  INTUITIVE  PARAMETERS 

The  proposed  models  will  be  developed  for  second  order  stationary  random  functions 
(Cressie  1991,  page  53).    Given  any  d  dimensional  covariance  function  C(h) 
corresponding    to    a    second    order    stationary    random    function  such 
that  J\C(h)  \  dh  <  °° ,  there  exists  a  d-dimensional  spectral  density  function  given 
by  the  Fourier  transform  (Bochner  1955) 


A  spectral  density  function  S(f)  is  isotropic  if  S(f)=G(  |  f  | )  for  some  scalar  function  G, 
and  radial  if  there  exists  some  d  dimensional  linear  transformation  A,  such  that 
S(f)=G(|  Af|).  In  other  words  a  space  with  radial  spectral  density  function  can  be 
deformed  through  linear  transformation  to  an  isotropic  space. 

We  propose  an  anisotropic  radial  spectral  model  with  an  ellipsoidal  base,  and  radial 
function  based  on  the  gamma  probability  density 


K 


(l) 


R 


P 


G»  =Knp"le 


(2) 


156 


where  p  represents  frequency  scaled  in  ellipsoidal  coordinates.  A  subset  of  the  family 
of  gamma  radial  functions  is  given  in  figure  1 . 

Intuitive  Motivation  of  Parameters 

At  large  scales,  many  physical  processes  exhibit  nearly-periodic  behavior  (i.e.  the 
human  eye  observes  repetition  although  the  pattern  also  has  a  random  component). 
Examples  include  topography  of  mountainous  terrain,  vertical  height  of  ocean  waves, 
recurrence  of  vegetative  patches  and 
porosity  of  fractured  media.  Simulations 
of  these  processes  at  large  scales  should 
include  these  "pseudo-periodicities". 
The  mode  frequency  of  the  gamma 
spectral  density  function  controls  the 
periodicity  of  a  simulation.  If  the 
spectrum  is  very  peaked,  (ce»l),  most 
of  the  variation  will  be  concentrated  in 
oscillations  at  a  single  frequency.  The 
simulation  will  be  nearly  perfectly 
periodic  like  an  egg  carton.  In  contrast, 
if  the  spectrum  is  very  flat  (ct~l),  the 
corresponding  simulated  data  will 
contain  oscillations  at  nearly  all 
frequencies  (white  noise).  Anisotropics 
controlled  by  the  ellipsoidal  base  of  a 
radial  spectrum  as  in  figure  (2)  cause  "ridges"  in  simulated  data.  Few  physical 
processes  are  isotropic  (one  rarely  sees  round  mountains)  and  hence  this  feature  is 
essential  to  realistic  simulations.  The  axes  of  the  ellipsoidal  base  are  easily  chosen  or 
estimated  from  data.  Finally  the  scaling  constant  K,,  may  be  chosen  to  give  simulated 
data  the  proper  variance.  Using  the  gamma  spectral  density  model  in  frequency 
domain,  these  properties  are  easily  controlled  to  produce  conditional  or  unconditional 
simulations.  The  same  properties  are  implicit  in  space  domain  methods  although  not 
dirrectly  controlled  by  the  practitioner. 

COVARIANCE  FUNCTIONS  CORRESPONDING  TO  RADIAL  SPECTRA 

Let  S(f)  be  a  radial  spectral  density  function  with  radial  function  G,  corresponding 
to  a  second  order  stationary  random  function  with  fe  Rn  (n  dimensional  Euclidean 
space),  and  ellipsoidal  base  given  by  fTBf  =  1,  where/?  =  VLVT,'is  the 
eigenvalue/eigenvector  decomposition  of  B.  Making  the  substitutions 
<t>  =  A12r7/;       \J\  =  \B\~m  and  h  =  VLmx,  in  equation  (1)  gives 


12: 

e 

10  i 

e 

08  -. 

c 

Spectrum  S 

a 
0 
a 

at  - 
at 

02 : 

0 

00  - 
0. 

00 

0.  20 

0.10  c.-,o 
Frequency  f 

0.  B0 

'  .  00 

Figure  1.  Family  of  radial  gamma  spectral 
models  with  mode  frequency  0.2  and  a 
ranging  from  1.5  to  5.5  and  P  =  l/(oc-l).  These 
models  all  have  dominant  wavelength  5  units 
with  varying  amounts  of  high  frequency  noise. 


157 


C(h)  =  jG{s[tf$)el2^\B\md$. 


Q 


(3) 


The  n-dimensional  integral  in  equation  (3),  can  be  reduced  to  a  one  dimensional 
Hankel  transform  through  introduction  of  the  hyper-spherical  coordinate 
transformation  (Miller  1964  ).  Computing  n-1  integrals  gives 


2     P  n~ 

CW  =  J  p~  G(p)  J(»-2>/2(* P)  v^P  d9,  <4> 

^  0 


where  R  -  TB  lh,  is  a  function  of 
h,  and  Jv  is  the  Bessel  function  of  order 
v.  This  is  the  Hankel  transform  of  order 
(n-2)/2  which  is  tabled  for  many 
common  functions  by  Ditkin  and 
Prudnikov  (1965  p.  432).  Note  that  C 
is  a  function  of  \L~V2VTh\  and  hence 
must  be  radial  with  an  ellipsoidal  base. 
Further,  the  covariance  function  is  given 
by  the  Hankel  transform  of  the  radial 

function        These  results  for  second     Figure  2.  Two  dimensional  radial  gamma 
order  stationary  random  functions  can    spectral  density  function  with  3  to  1 
be  combined  with  the  converse  from  anisotropy. 
Taheri  (1980)  and  Hagen  (1982). 

In  summary,  we  have  that  for  S(f)  and  C(h),  the  spectral  density/covariance  pair  with 
radial  functions  G  and  g  respectively;  S(f)  is  radial  with  ellipsoidal  base  fTBf  =  1  if 
and  only  if  C(h)  is  radial  with  ellipsoidal  base  h^^h  =  1,  and 

00 

C(h)  =  /  P("~1,/2G(P)4-2)/2(2*ptf)v/2^P  (5) 

0 


0  =  4.0    R*io  of  Anisotropy:  3:1 


if  and  only  if 

Sif)  =        /  R{n~xn  m)J{„-mV*pR)J^pRd<>.  (6) 

P  0 


158 


The  Gamma  Spectral  Model  for  n  Dimensional  Space 


These  results  can  be  used  to  derive  the  gamma  and  other  spectral  density 
covariance  pairs  in  n  dimensions.  Using  equation  (5)  and  the  Hankel  transform  of  the 
gamma  radial  function  (Ditkin  and  Prudnikov  1965,  p. 439,  eq.  11.38),  the  n 
dimensional  covariance  function  corresponding  to  the  gamma  spectral  model  is  given 
by 


n+2a 


ti-2 


C  (h)  =  K 


271  P 


!>+«  +  !) 


•p 


2\(^2a)/4  //+2a-2 


Ul+(2itprt)2) 


2\  1/2 


.  (7) 


where  P  (z),  is  the  Legendre  Spherical  function  (Stegun  p.  332,  eq.  8. 1 .2)  and  the 
scaling  constant  Kn  =  /W°M  2^a+n  l  r(a+/;-l)r(/?/2)7i"/2  )~\  is  obtained  by 
forcing  the  integral  of  the  spectrum  over 
the  full  frequency  space  to  be  o2.  A 
subset  of  the  class  of  radial  functions  is 
plotted  in  figure  (3).  It  should  also  be 
noted  that  the  class  of  covariance 
functions  is  quite  flexible  containing 
members  with  shapes  ranging  from  that 
of  the  Gaussian  to  the  hole  effect  model. 
Widely  varying  model  types  can  be 
chosen  through  parameter  estimation 
alone  as  opposed  to  fitting  several 
models  from  separate  classes  followed  by 
ad  hoc  procedures  for  choosing  the  best 
class  of  models. 

MULTI  VARIATE  CROSS- 
SPECTRA  IN  Rn 

To  simulate  multivariate  (vector)  random  processes  with  frequency  domain 
methods,  models  of  cross  spectra  are  required.  In  general  these  models  are  obtained 
through  modeling  the  cross  covariance  in  lag  domain.  This  is  sometimes 
unsatisfactory  in  that  the  cross  covariance  is  typically  not  symmetric  about  zero  and 
does  not  have  any  a-priori  expected  form.  However,  with  some  simplifying 
assumptions,  a  reasonable  cross  spectral  model  can  be  obtained  which  is  intuitive  in 
nature  In  particular,  it  is  assumed  that  the  coherence  and  phase  are  approximately 
constant  in  the  range  of  frequencies  with  significant  power  in  the  univariate  spectra. 

Let  x  be  an  n  dimensional  vector  in  K",  and  Vk(x),  k=l,2,3...d  be  a  set  of  d 


Figure  3.  Family  of  covariance  functions 
corresponding  to  the  gamma  spectral  model 
with  a  ranging  from  0.01  to  0.61  in  increments 
of  0.2  and  from  1  to  46  in  increments  of  5. 


159 


intercorrelated  second  order  stationary  random  functions.  The  cross  covariance 
function  between  the  ijth  pair  of  random  functions  is 


ctfh)  = 


(8) 


For  i  =  j,  this  is  the  auto  covariance  function.  The  cross  spectrum  for  any  pair  is 
■W)  =  [C^We-^dh  =  cff)  -%(f) 


where  c^  and  are  the  co-  and  quad-spectra.  In  general,  Q/h)  may  not  be  symmetric 
about  the  origin,  and  hence  S^f)  will  be  complex  valued.  However,  since  Cy(h)  is  real 
valued,  the  co-spectrum  is  even  and  the  quad-spectrum  is  odd  (Borgman,  1993).  The 
coherence  and  phase  are  defined  in  terms  of  the  co-  and  quad-spectra  as 


2         -  ^^(f) 

Slt(f)SM(f) 


*.,.(/)  =  arctan 


(10) 


These  relations  can  be  looked  at  as  a  polar  transformation  with  angle  4>y  and  radius 
r  =  yvylSaSr  Solving  for  Cy  and  qy  gives 

6n =  YyV^  cos$v  and  %  =  y.j^jj^u  (ii) 


where  the  argument  f  has  been  suppressed  (although  assumed  to  be  present)  for 
notational  convenience.  Note  that  if  is  zero,  then  the  co  and  quad-  spectra  are  also 
zero.  Therefore  values  for  Yy  and  are  required  only  for  those  frequencies  where 
the  individual  univariate  spectra  are  nonzero.  The  proposed  model  is 


stJ(f)  -  \ 


Y..^Af)SAf)e  'X    for    FL  <fTBf<  Fu  f>0 

V^ifliifle'*",    M    FL<fTBf<Fu  f<0 
0  elsewhere 


(12) 


where  the  phase  and  coherence  are  assumed  constant  in  the  range  of  frequencies 
where  the  univariate  spectra  are  non-zero,  and  where  FL  and  Fu  are  the  lower  and 
upper  bounds  on  that  region  The  model  is  defined  piecewise  to  allow  the  quad- 
spectrum  to  be  odd.  The  sets  on  the  right  side  of  equation  (12)  are  denoted  Q+  and 
Q"  respectively.  These  sets  could  be  generalized  to  any  half  space. 

To  apply  the  model,  estimates  of  S^,  Sy,  Yij,  and  are  required.  The  marginal 
spectra  may  be  estimated  and  modeled  independently  using  appropriate  spectral 


160 


models.  The  gamma  is  proposed  here  due  to  its  flexibility  in  shape  and  ease  of 
application,  although  other  spectral  models  may  be  used. 

Given  that  each  of  the  d  marginal  spectra  have  been  estimated  or  reasonably  chosen, 
the  coherence  and  phase  remain  to  be  estimated.  By  expressing  the  cross-covariance 
functions  in  terms  of  the  inverse  Fourier  transform  of  the  cross-spectra,  the  phase  and 
coherence  parameters  (assumed  constant)  may  be  estimated  in  lag  domain.  The 
Fourier  transform  of  the  cross  spectra  may  be  expressed  as  a  sum  of  sine  and  cosine 
integrals  C^h)  =  YyCOS^)  Ic(h)  +  Yvsin(4>;7)  ls(h),  where  Ic  and  Is  are  given  by 

Ic  ^^cos(2tt/^)#, 

Q 

and  (13) 

/  =  j  ^  sin(2n f  Th)  df  -  J/pVsin(27i/^)# 


If  CyCh)  has  been  estimated  possibly  with  some  form  of  bin  averaging  or 
nonparametric  smoothing  at  a  set  of  N  lags  {h*k,  k=l,2,3,...,N},  estimates  of  the 
phase  and  coherence  may  be  obtained  from  the  N  linear  equations  in  Y,yCos(|)f ,  and 

Y(/sin<tV 

Ct]{h\)  =  Y(/cos(t);//c(/i;)  +  y         Iflfo  =  g^^l)  +  gl]2Is{h\).  (14) 


These  equations  can  be  solved  by  least  squares  to  obtain  estimates  of  g^,  and  gjj2.  The 
estimated  phase  and  coherence  are  given  by 


~2 


fyj  =  arctan 


oij,  2 


An  example  of  a  cross  correlation 
function  corresponding  to  the  proposed 
gamma  model  is  plotted  in  figure  (4). 
Each  of  the  marginal  processes  have 
univariate  gamma  spectra  with  ellipsoidal 
bases.  The  maximum  correlation  is 
located  270°  counter  clockwise  from  the 
first  lag  domain  axis  hh  which 
corresponds  to  a  constant  phase  shift  of 
270°.  The  minimum  correlation  is 
negative  and  located  180°  from  the 
maximum  correlation  This  type  of  cross 


(15) 


Figure  4.  Cross  covariance  function  for  2 
correlated  anisotropic  random  functions  with 
radial  gamma  spectral  models,  coherence  1.0 
and  phase  angle  270°. 


161 


correlation  function  could  be  used  to  simulate  the  relation  between  rainfall  and 
surface  runoff  where  a  directional  lag  would  be  anticipated  in  the  spatial  correlations. 

DISCUSSION 

Frequency  domain  conditional  simulation  techniques  allow  the  user  to  directly 
control  pseudo-periodicities  anisotropy  and  high  frequency  noise  in  simulated  data 
sets.  The  univariate  gamma  spectral  density  model  allows  the  user  to  easily  control 
these  characteristics  by  directly  identifying  parameters  with  physical  properties.  The 
gamma  model  has  been  applied  by  and  Peacock  and  Kern  (1995)  to  model  hydraulic 
conductivity  in  the  Powder  river  basin  in  Wyoming  at  regional  scales. 

The  proposed  multivariate  models  provide  a  framework  for  estimating  cross  spectra 
or  equivalently  cross-correlation  functions  in  which  complicated  lag  domain  functions 
are  simplified  to  a  pair  of  band  limited  functions  of  frequency  which  are  typically  zero 
over  most  of  the  frequency  domain.  Further  research  into  this  area  should  include 
investigation  of  non-constant  functional  forms  for  the  phase  and  coherence. 

Frequency  domain  methods  should  be  of  particular  interest  to  geographers  and 
environmental  scientists  handling  large  data  sets  in  GIS  systems.  Large  scale 
simulations  at  computational  speeds  obtained  with  frequency  domain  methods  are  not 
feasible  with  space  domain  procedures. 

ACKNOWLEDGMENTS 

We  are  grateful  for  funding  of  this  work  which  came  from  a  cooperative  agreement 
between  the  state  of  Wyoming  Department  of  Environmental  Quality,  Wyoming  State 
Engineers  Office,  the  United  States  Office  of  Surface  Mines,  the  United  States  Bureau 
of  Land  Management  and  the  University  of  Wyoming. 

REFERENCES 

Bochner,  S.  1955.  Harmonic  Analysis  and  the  Theory  of Probability.  University  of 

California  Press,  Berkeley  and  Los  Angeles,  CA. 
Borgman,  L.E.  and  R.C.  Faucette,  1993.  Basic  mathematics  and  statistical  theory  for 

finite  Fourier  coefficients  of  Gaussian  vector  random  functions.  In: 

Computational  Stochastic  Mechanics,  Chap.  1,  Computational  Mechanics 

Publ.,  London. 

Borgman,  L.E.,  CD.  Miller,  S  R.  Signorini  and  R.C.  Faucette,  1994.  Stochastic 
interpolation  as  a  means  to  estimate  oceanic  fields.  Atmosphere-Ocean,  32(2) 
1994,  p.  395-419. 

Borgman  L.E.,  Taheri,  M.,  and  R.  Hagan,  1984.  Three-dimensional,  frequency- 
domain  simulations  of  geological  variables.   In  Geostatistics  for  Natural 
Resources  Characterization,  Part  7,  G.  Verly,  M.  David,  A  G.  Journel,  and 
A.  Marechal,  eds.  Reidel,  Dordrecht,  p.  517-541. 


162 


Cressie,  N.A.C.  1991.  Statistics  for  Spatial  Data,  John  Wiley  And  Sons,  Inc.  New 
York. 

Ditkin,  V.A.  and  A  P.  Prudnikov,  1965.    Integral  Transforms  and  Operational 
Calculus.  Translated  by  D.E.  Brown  and  edited  by  I.N.  Sneddon.  Pergamon 
Press,  New  York. 

Hagan,  R.  1982.  Application  of  spectral  theory  and  analysis  in  mining  geostatistics 
and  statistical  linear  wave  theory.  PhD.  Dissertation,  Department  of 
statistics,    University  of  Wyoming,  Laramie  WY. 

Journel,  A.  and  C.J.  Huijbregts,  1978.  Mining  Geostatistics,  Academic  Press,  New 
York. 

Journel,  A.  and  E  H.  Isaaks.  1984.    Conditional  simulation:  applications  to  a 
Saskatchewan  uranium  deposit.  Mathematical  Geology,  16:685-7 18. 

Krige,  D.G.   1951.  A  statistical  approach  to  some  mine  valuaation  and  allied 
problems  at  the  Witwaterstrand.  Unpublished  masters  thesis,  University  of 
Witwaterstrand. 

Matern,  B.  1986.  Spatial  Variation,  In  Lecture  Notes  in  Statistics  No  36,  Edited  by: 
D.  Brillinger,  S.  Fienberg,  J.  Gani,  J.  Hartigan  and  K.  Krickeberg,  Springer 
Verlag,  NewYork. 

Matheron,  G.  1971,  The  theory  of  regionalized  variables  and  it's  application.  Re.  No. 
5,  Les  Cashiers  de  centre  de  morphologie  mathematique  de  fontaingleau,  C.G. 
Fontainbleau. 

Peacock,  K.  and  J.  W.  Kern.  1995.  Assessment  of  groundwater  impacts  related  to 
the  proposed  Lighthouse  coal  bed  methane  project.  United  States 
Department  of  Interior,  Bureau  of  Land  Management,  Casper  District. 
Casper  WY. 

Stegun,  I.  A.  1972.  Legendre  functions,  In:  Handbook  of  Mathematical  Functions 

with  Formulas,  Graphs  and  Mathematical  Tables.  Edited  by  M.  Abramowitz 

and  I. A.  Stegun.  p  331-354. 
Taheri,  S.  M.,  1980.  Data  Retrieval  and  Multidimensional  Simulation  of  Mineral 

Resources.    PhD  Dissertation,  University  of  Wyoming,  Department  of 

Statistics,    Laramie  Wyoming. 

BIOGRAPHICAL  SKETCH 

Leon  E.  Borgman  is  a  professor  of  statistics  and  geology  and  geophysics  at  the 
University  of  Wyoming,  Laramie,  Wyoming.  He  has  specialized  in  application  of 
stationary  random  function  theory  to  a  wide  range  of  environmental  science, 
hydrgeologic  and  oceanographic/civil  engineering  problems. 

John  W.  Kern  is  a  consulting  statistician  with  Western  Ecosystems  Technology  Inc., 
Cheyenne,  Wyoming.  John  has  primarily  specialized  in  application  of  stationary 
random  function  theory  to  environmental,  hydrogeological  and  ecological  problems. 


163 


Computing  the  Area  Affected  by 
Phosphorus  Runoff  in  an  Everglades 
Wetland  Using  Bayesian  and 
Universal  Kriging 

Song  S.  Qian 


Abstract. --Phosphorus-enriched  agriculture  runoff  is  believed  to  be  the  leading  cause 
of  ecosystem  changes  in  an  Everglades  wetland.  In  order  to  study  the  effects  of  the 
added  nutrients  to  the  wetland  ecosystem,  it  is  necessary  to  estimate  the  acreage  of  the 
affected  region.  In  this  study,  Bayesian  and  universal  kriging  are  used  to  analyze  the 
data  collected  by  Reddy,  et  al  (1991).  The  background  level  of  the  STP  concentration  is 
used  as  an  indicator  of  whether  the  region  is  affected  or  not  through  an  indicator 
function.  The  expected  value  of  the  affected  area  is  calculated  using  the  results  from 
Bayesian  and  universal  kriging.  The  results  show  that  universal  kriging  is  sensitive  to 
the  specification  of  covariance  function.  Universal  and  Bayesian  kriging  yield 
comparable  results  when  the  covariance  functions  are  specified  in  the  like  manner. 


INTRODUCTION 

The  Water  Conservation  Area-2A  (WCA2A)  is  part  of  the  Everglades  wetlands  in 
south  Florida  (Figure  1).  It  receives  the  agricultural  runoff  from  the  Everglades 
Agriculture  Area.  The  phosphorus  enriched  agriculture  runoff  caused  some  significant 
changes  in  this  phosphorus  limited  wetland  ecosystem.  The  most  obvious  change  was  in 
the  change  of  the  dominant  vegetation  species  from  sawgrass  to  cattail  near  the  inlet  of  the 
runoff.  The  significant  increase  of  phosphorus  level  in  the  water  of  WCA2A  is  considered 
a  major  threat  to  the  Everglades  National  Park.  To  protect  the  Park,  the  South  Florida 
Water  Management  District  proposed  to  use  constructed  wetlands,  a  man-made  marsh  like 
buffer  area,  to  remove  excess  phosphorus  before  the  water  enters  WCA2A. 

One  important  design  parameter  of  a  constructed  wetland  is  the  unit  mass  loading  rate. 
The  unit  area  loading  rate  is  usually  calculated  by  measuring  the  influent  mass  loading  rate 
(in  mass  of  phosphorus  per  unit  time)  and  dividing  it  by  the  area  of  the  receiving  wetland. 
Because  WCA2A  covers  such  a  large  region,  not  all  parts  of  the  region  are  effective  in 
removing  phosphorus.  (In  other  words,  not  all  parts  are  affected.)  If  the  entire  WCA2A 
were  taken  as  affected  and  used  to  calculate  the  mass  loading  rate,  the  phosphorus 
assimilating  capacity  of  the  wetland  is  underestimated. 

The  data  used  in  this  paper  are  from  Reddy  et  al.  (1991),  who  sampled  soils  at  74 
stations  on  seven  north  —  south  transacts,  spaced  at  2  mile  intervals  (Figure  2).  Richardson 
(personal  communication,  1993)  believes  that  the  STP  content  should  be  fairly  stable  if 
there  were  no  agricultural  runoff  problem,  and  the  background  level  of  the  STP  content  is 


164 


Figure  1.  Map  of  south  Florida  and  location  of  the  study  region 


Figure  2.  Locations  of  sampling  sites  in  WCA2A 


165 


about  500  (ig  of  phosphorus  per  gram  of  soil  (fig/g)  in  the  top  20  cm.  Universal  kriging 
(Cressie,  1991)  and  Bayesian  kriging  (Handcock  and  Stein,  1993)  are  used  to  model  the 
phosphorus  concentration  over  the  region.  When  the  concentrations  are  larger  than  the 
background  level,  it  is  believed  that  the  corresponding  region  is  affected  by  the  agricultural 
runoff. 

METHODS 

Data  analysis 

The  mean  STP  content  in  the  top  20  cm  layer  of  the  Reddy  samples  were  used.  The 
STP  values  are  log-transformed  to  stabilize  the  variance.  The  spatial  coordinates  were 
converted  from  latitude  and  longitude  to  the  Universal  Transverse  Mercator  (UTM)  grid 
system 

Universal  kriging  (UK)  can  be  interpreted  as  a  Gaussian  random  field  model: 

Z  =  Xp  +  Ti  (1) 
where:  Z  is  the  log  transformed  STP  content,  X  is  the  known  design  matrix 
containing  the  longitude  and  latitude,  and  the  distance  to  the  nearest  pollution 
source,  P  is  a  vector  of  unknown  parameters,  and  r|  is  the  error  term  which  is 
assumed  to  be  normally  distributed  with  mean  0  and  covariance  Z.  The  covariance 
function  is  cov{Z{sx),  Z(s2)}  =  aKQ(sv  s2),  for  any  pair  of  spatial  coordinate  points 
0,,  s2)  of  interest,  where  a  >  0  is  a  scale  parameter,  and  9  is  a  vector  of  structural 
parameters  which  specify  the  shape  of  the  covariance  function.  In  the  case  of 
kriging,  we  observe  {Z{sx),  Z(s2),  Z(sn)}'  =  Z  and  the  prediction  of  Z(s0)  is  the 
objective,  where  s0  is  a  location  with  no  observed  data.  The  UK  predictor  is  the 
best  linear  unbiased  predictor  (Cressie,  1991),  of  the  form  Ze(s0)  =  X(Q)'Z.  See 
Ripley  (1981)  for  details. 

The  covariance  matrix  aKe  is  usually  unknown,  and  is  often  estimated  using 
the  empirical  semi-variogram.  Once  the  semi-variogram  model  is  chosen,  the 
covariance  function  is  taken  as  known,  and  the  uncertainty  in  the  selected  model  is 
ignored.  To  include  this  uncertainty  in  the  analysis,  Handcock  and  Stein  (1993) 
introduced  Bayesian  kriging  (BK)  using  the  Matern  class  of  covariance  functions, 
with  a  parameter  0  =  (0,,  92). 

The  posterior  predictive  distribution  for  Z(sQ)  is  derived  by  adapting  a 
noninformative  prior: 

pr(a,  p,  0)ocpr(0)/a  (2) 
Handcock  and  Stein  (1993)  showed  that  the  conditional  posterior  distribution 
of  Z(s0)  is: 

ZCgie,  Z -  f„.,(z6(*0),  ^a(6) V0(8))  (3) 

where  q  is  the  number  of  parameters  in  (3. 
The  marginal  posterior  distribution  of  9  is: 

pr(Q\Z)  oc  pr(9)l^er,/2IXXe-1XI  ,/2a(9)-(',-9)/2  (4) 


166 


The  Bayesian  predictive  distribution  for  Z(sQ)  is: 

pr(Z(sQ)\Z)  oc  J  pr(Z(s0M  Z)  x  pr(Q\Z)dQ  (5) 


-i, 


where:  ZQ(s0)  is  the  UK  point  predictor,  a(6)  =  ~(Z  -  X$)XKQ\Z  -  Xp),  aV0(9)  is 
the  UK  prediction  error  variance,  and  t  is  a  student-/  distribution  with  (n  -  q) 
degrees  of  freedom.  See  Handcock  and  Stein  (1993)  for  details. 

The  prior  distribution  of  the  parameters  of  the  covariance  function  was  chosen 

as: 

^e)  =  Me,)xMej  =  (|+B|)'(1+ei)1  (6) 

This  prior  distribution  represents  the  belief  that  large  values  of  the  parameters 
are  less  likely  than  small  ones. 

Area  estimation 

The  expected  value  of  the  area  can  be  estimated  by: 

E{Ap)  =  j[j  I(Z(X)  >  t)dX]  x  pr(Z(X)  I  Z)dZ  =  J  pr(Z(X)  >  1 1  Z)dX  (7) 
where  /(•)  is  an  indicator  function,  X  is  the  coordinate  vector,  and  t  is  the 
background  level. 

When  UK  is  used,  the  integrand  in  the  RHS  of  equation  (7)  can  be  evaluated 
by  using  the  CDF  of  a  normal  density.  When  BK  is  used: 

E(Ap)=  J  J  pr(Z(X)  >/I6,Z)x  pr(Q  I  Z)dQdX  (8) 

XeR2  0 

The  uncertainty  of  the  estimated  area  using  BK  is  evaluated  using  a  Monte 
Carlo  simulation  method.  The  procedure  is  based  on  the  fact  that  given  9,  the  area 
is  Ap  =  \l{Z{X)  >  t\Z,  Q)dX.  It  can  be  approximated  by  sampling  Z(X)  and 
evaluating  the  integrand,  and  the  mean  of  these  values  is  approximately  equal  to 
the  proportion  of  the  affected  area  over  the  entire  WCA2A.  A  Monte  Carlo 
simulation  algorithm  was  used  to  sample  6  and  then  Z(X). 


RESULTS 


The  residuals  of  the  model  (1)  are  found  to  be  intrinsically  stationary  and 
isotropic  through  visual  inspection  of  the  semi-variograms  in  the  north-south  and 
east-west  directions.  For  UK,  four  different  covariance  functions  are  used.  One  is 
the  Matern  covariance  function  using  the  maximum  posterior  estimate  of  6 
(UK-M).  The  other  are:  exponential  (UK-exp),  spherical  (UK-sph),  and  Gaussian 
(UK-Gau).  Figure  3  shows  the  log-transformed  data. 

Covariance  function 

The  covariance  function  used  in  BK  is  presented  in  Figure  4,  the  log 
transformed  posterior  as  a  function  of  the  parameters  (6,  and  62).  The  posterior 


167 


550        555        560        565        570  2.5       3.0       3.5       4.0       4.5       5.0  5.5 

Eastinf  (km)  The  range  parameter 

Firure  3.  Log-transformed  data  Figure  4.  Posterior  of  the  covariance  function  parameters 


168 


mode  is  at  (3.6515,  -0.4697),  or,  the  Maximum  Posterior  Estimates  (MPE)  of  the 
parameters  are  (4482,  0.3391).  The  estimated  a  given  the  MPE  of  9  is  0.032363. 

The  covariance  functions  in  UK  are  estimated  through  fitting  the 
semi-variogram  model.  Figure  5  compares  the  covariance  functions  computed 
from  different  semi-variogram  models.  In  Figure  5,  the  line  labeled  "Matern"  is  the 
Matern  covariance  function  using  the  MPE  of  6. 

Predictive  distributions 

The  predictive  distributions  of  the  log-transformed  STP  are  estimated  for  the 
10th  sampling  site,  which  has  the  largest  STP  content,  and  the  25th  sampling  site, 
which  has  a  below  average  STP  content.  The  density  functions  are  shown  in 
Figures  6  and  7.  Table  1  compares  the  results  from  BK  and  UK  using  the  three 
semi-variogram  models. 


Table  1.  Predictive  distributions 


Model 

Sampling  Site 

Predicted  Mean 

Prob.  >  500  ng/g 

Measured  STP 

BK 

25 

2.48754 

0.076779 

2.398536 

UK-M 

25 

2.478177 

0.052068 

2.398536 

UK-exp 

25 

2.462711 

0.030408 

2.398536 

UK-Gau 

25 

2.179858 

0.000001 

2.398536 

UK-sph 

25 

2453698 

0.022335 

2.398536 

BK 

10 

2.826391 

0.786981 

3.100039 

UK-M 

10 

2.850576 

0.848553 

3.100039 

UK-exp 

10 

2.83751 

0.823649 

3.100039 

UK-Gau 

10 

2.814715 

0.736275 

3.100039 

UK-sph 

10 

2.843363 

0.835841 

3.100039 

Comparing  the  five  predictive  distributions  for  the  10th  sampling  site,  we  note 
that  the  predictive  distributions  from  UK-exp,  UK-sph,  and  UK-Mat  are  nearly 
identical.  BK  yields  a  similar  mean  value  but  slightly  larger  estimation  variance. 
The  predictive  distribution  from  UK-Gau  has  the  largest  variance.  For  the  25th 
sampling  site,  we  see  comparable  predictive  distributions  from  UK-exp,  UK-sph, 
and  UK-M.  BK  yields  a  similar  mean  and  slightly  larger  variance.  UK-Gau 
produces  a  much  different  predictive  distribution  from  the  other  four.  The  mode  is 
significantly  less  then  the  modes  from  the  other  four  predictive  distributions.  The 
Gaussian  covariance  function  is  equivalent  to  the  Matern  covariance  function  with 
9;  — >  oo.  This  means  that  UK-Gau  describes  a  function  which  is  infinitely 
differentiable.  This  representation  may  not  be  realistic. 

Computation  of  the  affected  area 

Equation  (6)  was  used  for  UK  and  Equation  (7)  was  used  for  BK.  The  results 
are  listed  in  Table  2. 

In  this  study,  we  did  not  try  to  locate  the  boundary  of  the  affected  region. 
However,  the  shape  of  the  affected  region  may  be  stable  and  we  plot  the  contour 


169 


lines  of  the  probability  of  STP  larger  than  500  |j,g/g  to  indicate  the  possible  shape. 
Figure  8  is  plotted  using  BK,  and  Figures  9  to  12  show  the  surfaces  from  UK.  The 
surface  from  the  UK-Gau  is  mostly  flat  but  has  sudden  jumps.  The  reason  for  this 
unnaturally  shaped  surface  is  that  the  Gaussian  covariance  function  describes 
"super  smooth"  surfaces,  in  this  case,  super  smooth  for  the  residuals.  A  model 
with  very  smoothed  residual  surface  must  have  a  jumpy  mean  surface. 


Table  2.  Expected  area 


Model 

affected  area  (km2) 

%  of  the  total  area 

BK 

97.186 

22.394 

UK-M 

95.775 

22.069 

UK-sph 

89.403 

20.601 

UK-exp 

99.615 

22.954 

UK-Gau 

76.588 

17.648 

From  Table  2,  we  note  that  the  estimated  area  using  BK  is  very  close  to  the 
estimated  area  using  UK-M  and  UK-exp.  The  estimated  area  is  sensitive  to  the 
specification  of  the  covariance  function  when  UK  is  used.  It  is  reasonable  to  state 
BK  is  more  appropriate  in  this  study,  since  the  covariance  function  is  unknown  and 
the  uncertainty  in  estimating  the  covariance  function  is  considered. 

The  uncertainty  of  the  estimated  area  is  evaluated  only  for  BK.  Figure  13 
presents  the  histogram  of  the  percentages  of  the  affected  area  from  the  Monte 
Carlo  simulation. 

DISCUSSION 

In  conclusion,  it  is  seen  that  one  significant  difference  is  an  increased 
computational  intensity  of  BK.  However,  the  results  indicate  that  UK  is  sensitive 
to  the  specification  of  the  covariance  function.  The  difference  between  two 
different  covariance  functions  may  be  significant  even  though  the  semi-variogram 
models  of  the  two  are  fitted  to  the  data  equally  well.  This  sensitivity  to  the 
covariance  function  justifies  the  use  of  BK. 

There  are  several  studies  that  estimated  the  phosphorus  affected  area  of 
WCA2A.  Craft  and  Richardson  (1993)  use  the  same  data  set  as  presented  in  this 
paper.  In  their  study,  the  boundary  of  the  affected  region  was  delineated  by  using 
both  the  STP  content  and  the  phosphorus  accumulation  rate.  The  delineated 
boundary  is  plotted  onto  a  USGS  topographic  map,  and  the  area  is  thus  measured. 
Their  estimate  of  the  area  is  115  km2.  Only  the  top  10  cm  layer  of  the  Reddy  data 
was  used  and  600  jLLg/g  is  the  background  level.  The  STP  contour  lines  were 
produced  using  the  ordinary  kriging  algorithm  from  a  commercial  graphic  software 
(Surfer®).  The  semi-variogram  model  used  is  not  reported. 

Using  the  STP  content  as  an  indicator  of  the  influence  of  the  agricultural  runoff 
is  a  more  appropriate  way  of  estimating  the  affected  area.  A  spatial  statistical 


170 


171 


approach  accounts  for  the  uncertainty  in  the  sample  data.  Considering  the  high 
level  of  uncertainty  involved  in  the  estimating  process  (Figure  13),  the  estimated 
area  using  BK  should  be  considered  comparable  to  the  result  in  Craft  and 
Richardson  (1993).  The  discrepancy  between  the  estimates  may  be  caused  by  two 
factors,  without  considering  the  error  introduced  by  the  process  of  measuring  the 
area  from  a  map.  They  are  their  use  of  ordinary  kriging  and  may  be  a  different 
covariance  function,  and  the  different  background  level.  There  is  no  reason  to 
believe  that  the  background  level  is  a  constant.  The  field  samples  from  the  region 
that  is  unaffected  by  the  agriculture  runoff  show  a  significant  variation  (Craft  and 
Richardson,  1993).  In  a  current  study  by  the  author,  the  background  level  is 
represented  by  a  probability  distribution. 

ACKNOWLEDGMENT 

Drs.  Richard  Smith  and  Peter  Mtiller  motivated  the  author  to  pursue  this 
project.  Dr.  Marcia  Gumpertz's  comments  and  suggestions  significantly  improved 
this  article.  The  author  appreciate  the  encouragement  and  support  from  Drs.  C.J. 
Richardson  and  K.H.  Reckhow. 

REFERENCES 

Craft,  C.B.  and  C.J.  Richardson.  1993.  Peat  accretion  and  phosphorus 

accumulation  along  a  eutrophication  gradient  in  the  northern  Everglades. 
Bio  geochemistry,  22,  133-156. 

Cressie,  N.A.C.,  1991.  Statistics  for  Spatial  Data,  John  Wiley  &  Sons,  Inc.  New 
York. 

Handcock,  M.S.,  and  M.L.  Stein.  1993.  A  Bayesian  analysis  of  kriging, 

Technometrics,  35(4),  403-410. 
Ho  ward- Williams,  C.  1985.  Cycling  and  retention  of  nitrogen  and  phosphorus  in 

wetlands:  a  theoretical  and  applied  perspective,  Freshwater  Biology,  15, 

391-431. 

Reddy,  K.R.,  W.F.  DeBusk,  Y.  Wang,  R.  DeLaune,  and  M.  Koch,  1991. 

Physico-chemical  properties  of  soils  in  the  Water  Conservation  Area  2  of  the 
Everglades.  Final  report  submitted  to  South  Florida  Water  Management 
District,  West  Palm  Beach,  FL  33416. 

Ripley,  B.D.  1981.  Spatial  Statistics,  John  Wiley  &  Sons,  Inc.  New  York. 

BIOGRAPHICAL  SKETCH 

Song  S.  Qian  is  a  postdoctoral  research  associate  at  the  Duke  Wetland  Center  of  the 
Nicholas  School  of  the  Environment  of  Duke  University.  He  received  his  Ph.D.  in 
environmental  sciences  and  MS  degree  in  statistics  from  Duke  University.  He  has  a 
Master's  degree  in  environmental  systems  engineering  from  Nanjing  University,  China. 


172 


Mapping  Synecological  Coordinates: 
A  Spatial  Analysis  of  Environmental 
Indices  in  a  Forested  Landscape 

Margaret  R.  Holdaway1  and  Gary  J.  Brand2 


Abstract — The  Method  of  Synecological  Coordinates  was  used  to 
compute  environmental  indices  (synecological  coordinates)  for 
moisture  and  nutrients  based  on  plant  species  presence.  Plots  were 
sampled  in  a  65.6-ha  area  in  St.  Louis  County,  Minnesota,  dominated 
by  quaking  aspen.  A  four-level,  nested  sampling  design  quantified 
spatial  variability  over  a  wide  range  of  distances.  Semivariances  at 
the  four  distances  suggested  a  45-m  equidistant  grid  of  plots  would 
be  appropriate  for  mapping  synecological  coordinates.  Variograms 
of  moisture  and  nutrients  coordinates  showed  ranges  of  350  m  or 
greater,  indicating  that  spatial  dependence  continues  well  beyond  the 
limits  normally  found  with  most  soil  studies.  Nutrients  coordinate 
showed  greater  spatial  dependence  than  moisture  coordinate,  which 
showed  more  random  variation.  The  nutrients  coordinate  map  proved 
useful  in  defining  areas  of  differing  quaking  aspen  site  index.  Thus, 
the  efficient  mapping  of  site  synecological  coordinates  has  helped 
capture  the  pattern  of  environmental  variation.  


INTRODUCTION 

Our  ability  to  appropriately  manage  a  forest  community  can  be  improved  if  we 
can  accurately  map  environmental  conditions.  We  propose  a  general  method  that 
maps  environmental  indices  (plant-based  indicators).  The  intent  of  this  paper  is  to 
investigate  whether  environmental  indices  can  be  accurately  mapped,  and  if  so, 
whether  the  maps  identify  areas  of  contrasting  site  productivity. 

Environmental  indices  are  produced  by  the  Method  of  Synecological 
Coordinates  (MSC)  (Bakuzis  1959,  Bakuzis  and  Kurmis  1978,  Brand  1985,  and 
Gutierrez-Espeleta  1991).  MSC  uses  plant  presence  as  an  indicator  of  the 
environment.  It  is  based  on  the  notion  that  floristic  composition  of  a  site  indicates 
the  nature  of  the  environment  there.  MSC  is  an  easily  applied  approach  that 
computes  semi-quantitative  values  for  four  environmental  factors — moisture  (M), 
nutrients  (N),  heat  (H),  and  light  (L).  M  and  N  represent  the  edaphic  conditions 
and  H  and  L  represent  the  climatic  conditions  at  the  site.  Each  plant  species  has  its 
own  set  of  synecological  coordinate  values  on  a  relative  scale  of  from  1  to  5, 
representing  on  average,  the  species  "need"  for  that  factor.  An  arithmetic  mean  of 
the  coordinates  for  those  species  occurring  together  is  called  the  site  synecological 


*  Mathematical  Statistician,  North  Central  Forest  Experiment  Stn,  St.  Paul.  MN. 
-  Research  Forester.  North  Central  Forest  Experiment  Stn,  St.  Paul,  MN. 


173 


coordinate  for  that  factor.  For  this  paper  we  will  concentrate  on  edaphic 
environmental  conditions  as  represented  by  M  and  N. 

Synecological  coordinates,  which  use  the  integrative  nature  of  plants  and  prior 
knowledge  of  their  environmental  "requirements,"  may  provide  a  practical  means 
to  determine  the  spatial  variability  and  distribution  of  environmental  factors. 
Spatial  variability  is  analyzed  using  geostatistical  techniques.  Geostatistical  theory 
is  based  on  the  simple  notion  that  measurements  made  closer  together  in  space  are 
more  alike  than  those  made  farther  apart.  The  spatial  variability  or  structure  can 
be  modeled  via  a  mathematical  function  known  as  the  variogram.  Once  the 
variogram  is  modeled,  kriging  uses  information  in  the  variogram  to  spatially 
interpolate  unknown  data  points. 

The  spatial  variation  in  soil  properties  has  been  extensively  analyzed  with  geo- 
statistical techniques  (Oliver  and  Webster  1991,  Burgess  and  Webster  1980,  and 
Webster  and  Burgess  1983).  Because  soil  is  such  an  important  determinant  of  the 
environmental  conditions  of  a  site,  describing  the  spatial  distribution  of  soil 
properties  should  spatially  describe  important  aspects  of  the  environment. 
However,  the  chemical  and  physical  composition  of  soils  is  complex  and 
extremely  variable  over  short  distances  (Lechowicz  and  Bell  1991,  Jackson  and 
Caldwell  1993,  Oliver  and  Webster  1991).  Therefore,  synecological  coordinates 
may  provide  a  more  effective  means  to  integrate  soil  variability  and  complexity 
into  maps  of  edaphic  conditions. 

The  choice  of  sampling  distance  to  use  in  any  study  depends  on  the  range  over 
which  the  phenomena  are  spatially  correlated.  Complex  multi-step  designs  are 
especially  suited  for  exploring  the  full  spatial  structure  when  there  is  no  previous 
knowledge  of  the  spatial  nature  of  the  data.  Designs  containing  several  sampling 
scales,  such  as  a  systematic-cluster  design  (Fortin  et  al.  1989)  or  a  nested  design 
(Oliver  and  Webster  1986)  are  ideally  suited  for  such  investigations.  Sampling  is 
performed  in  two  stages.  In  the  initial  stage,  a  complex  sampling  scheme  is 
carried  out  over  a  wide  range  of  distances  to  ascertain  the  range  of  the  process. 
This  information  then  helps  establish  the  sampling  grid  in  the  second  stage,  where 
points  are  sampled  on  an  equal-spaced  grid  at  a  scale  well  within  the  spatial  range. 

We  are  investigating  synecological  coordinates  and  geostatistics  as  methods  to 
represent  the  local  edaphic  environment  in  an  aspen-dominated  forest.  By 
representing  the  edaphic  environment,  the  medium  of  forest  growth,  we  expect  to 
make  inferences  about  overstory  site  productivity.  Our  purpose  here  is:  (1)  to 
analyze  the  spatial  structure  and  scale  of  spatial  variability  of  site  synecological 
coordinates  for  moisture  and  nutrients  using  geostatistical  techniques,  (2)  to  assess 
the  accuracy  of  site  coordinates  mapped  by  kriging,  and  (3)  to  examine  what  the 
mapped  site  coordinates  indicate  about  site  productivity. 

METHODS 

The  stage  1  study  site  (larger  square  in  figure  1)  covers  65.6  ha  in  St.  Louis 
County,  30  km  north  of  Duluth,  Minnesota.  A  timber  inventory  (based  on  limited 
field  sampling)  delineated  stand  boundaries  and  measured  site  index  (SI),  the 
traditional  measure  of  site  quality  in  forest  management.  Quaking  aspen  is  the 
primary  overstory  species.  The  topography  is  gently  rolling  with  elevations 
ranging  from  about  1440  feet  to  slightly  more  than  1490  feet  with  no  obvious 
directional  trends  in  the  topography. 


174 


Figure  1. — Aspen-dominated  landscape  in  St.  Louis  County,  Minnesota.  Heavy  line  shows 
road,  light  lines  delineate  stand  boundaries,  and  squares  delineate  study  areas.  Shading 
shows  the  site  index  class  (SI)  of  the  aspen  stands. 

A  nested  sampling  design  was  selected  to  provide  variability  data  for  a  wide 
range  of  distances.  In  stage  1  of  the  sampling  design,  we  determined  the 
variability  at  coarse  (135  m),  moderate  (45m),  fine  (15  m)  and  very  fine  (5m) 
distances.  The  sampling  distances  needed  to  be  simultaneously  small  enough  to 
estimate  the  nugget  variance  (unexplained  or  random  variance  that  occurs  at  very 
small  sampling  distances)  and  large  enough  to  identify  the  spatial  range. 

In  stage  1,  the  primary  or  coarsest  grid  of  points  (level  1)  had  separation 
distances  of  135  m.  The  axes  of  the  sampling  grid  were  oriented  in  a  north-south, 
east-west  direction  with  the  starting  grid  point  randomly  located.  A  geometric 
progression  of  decreasing  distances  between  sample  points  was  used.  Therefore, 
the  distances  for  the  second-level  points  were  45  m.  Two  second-level  points  were 
located  45  m  to  the  south  and  west  of  each  first-level  point  (see  figure  2a).  Each 
primary  and  secondary  point  included  third-  and  fourth-level  points  (figure  2b). 
The  two  third-level  points  were  located  15  m  to  the  south  and  west  of  every 
primary  and  secondary  point;  the  two  fourth-level  points  were  located  5  m  to  the 
north  and  east  of  every  primary  and  secondary  point. 

Collecting  plant  lists  at  each  point  in  a  complete  four-level  sampling  design 
would  have  been  too  intensive  for  the  time  and  money  available.  Therefore, 
alternating  primary  rows  and  columns  were  dropped  from  the  study  for  second- 
level  points  (figure  2a)  and  the  number  of  third-  and  fourth-level  points  were 
reduced  by  randomly  selecting  half  of  the  level  1  and  level  2  points  for  sampling 
at  levels  3  and  4.  At  each  point,  plant  species  occurring  within  a  square  plot  of 
5x5  m  were  recorded  and  synecological  coordinates  were  computed  using  a  data 
recorder  and  program  (Nimerfro  and  Brand  1993).  Species  synecological 
coordinates  developed  for  Minnesota  forests  were  used  in  the  computations 
(Baku/is  and  Kurmis  1978). 

Obvious  outliers  were  removed  from  the  data  at  each  stage.  A  plot  was 
removed  if  its  value  for  either  factor  was  judged  an  outlier  by  the  box  plot.  The 
stage  1  semivariance  based  on  sampling  at  four  lag  distances  (5,  15,  45,  and  135 


175 


o       o       o       o       o      e  o 


Figure  2. — Final  sampling  design:  (a)  coarse  and  moderate  distances,  (b)  fine  and  very 
fine  distances. 


m)  was  plotted  against  distance.  Ranges  were  estimated,  and  an  appropriate 
sampling  distance  was  chosen  from  among  the  four  lag  distances  for  stage  2  plot 
spacing. 

In  stage  2,  the  experimental  unidirectional  variograms  were  calculated.  Cross 
validation  was  performed  to  check  the  validity  of  the  final  models.  Differences 
between  kriged  estimated  and  observed  values  were  analyzed  using  these  cross 
validation  statistics:  mean  error,  median  absolute  error,  and  MSE. 

VARIOGRAM  MODELING  AND  CROSS  VALIDATION 

Site  synecological  coordinates  were  computed  from  254  plots  at  stage  1  grid 
points.  At  an  additional  13  plots,  plant  lists  could  not  be  collected  due  to 
environmental  conditions  detrimental  to  terrestrial  vegetation  (road,  permanent 
standing  water,  etc.)  or  site  coordinates  were  outliers.  The  semivariances  at  the 
four  lag  distances  for  stage  1  data  (black  dots  on  figure  3)  show  increasing  spatial 
variability  at  135  m  for  N.  However,  this  distance  may  be  too  large  when  kriging 
M,  so  we  chose  45  m  as  the  grid  sampling  distance  for  the  stage  2  analysis. 

Stage  2  plots  (smaller  square  in  figure  1)  were  located  on  a  1  lxl  1  grid  centered 
within  the  stage  1  area.  Nine  plots  out  of  121  were  located  on  the  road  or  in 
permanently  wet  areas  and  were  ignored.  An  additional  five  plots  were  removed 
because  their  site  synecological  coordinates  were  outliers.  A  summary  of  site 
coordinates  for  the  remaining  107  plots  shows  similar  means  for  M  and  N  of  just 
under  2.5  with  values  ranging  roughly  from  2.0  to  3.0.  Nutrients  had  slightly 
greater  range  and  standard  deviation  than  moisture. 

Isotropic  variograms  were  calculated,  using  a  minimum  distance  of  45  m  and  a 
maximum  of  500  m  (75%  of  the  largest  distance.)  Models  were  fit  to  the 
experimental  variograms  (figure  3).  Model  and  parameters  were  estimated 
visually  and  provided  excellent  fits  with  R2  of  .943  for  M  and  .985  for  N.  The 
linear  model  was  fit  to  M,  the  spherical  model  to  N.  The  experimental  variogram 


176 


values  for  stage  2  N  data  progressively  increased  with  distance  and  then  showed  a 
systematic  decrease,  known  as  a  hole  effect,  beyond  the  range  of  spatial 
dependence.  Since  variograms  were  plotted  beyond  half  the  maximum  separation 
distance,  the  apparent  hole  effect  is  likely  an  edge  effect.  The  models  are  fitted  for 
values  up  to  the  range.  In  addition,  anisotropy  was  examined.  However,  there 
were  not  enough  data  to  accurately  estimate  apparently  mild  directional 
influences. 


0.03  - 


o  0  02 

o 
c 

.00 
k_ 
03 

> 
E 

CD 

CO  0.01 


0.00 


a)  IVI 

<a  a 

• 

1 

Stage  1 

• 

Stage  2 

□ 

1 — f 


0       100    200    300    400  500 

Distance  (m) 


0.06  - 


0.05  - 


0.04 


0.03  - 


0.02- 


0.01  - 


0.00 


b)  N 

ll/i  I  l — 1 

□ 

□ 

□ 

/□ 

/a 

— 1 — I 


0       100    200    300    400  500 
Distance  (m) 


Figure  3. — Experimental  and  fitted  variograms  for  site  synecological  coordinates:  (a) 
Moisture  and  (b)  Nutrients. 


The  supplemental  information  about  the  spatial  variability  at  close  distances 
confirms  that  there  is  at  most  a  moderate  nugget  effect  (i.e.,  random  variation  or 
spatial  dependence  at  less  than  5  m)  in  the  experimental  variograms,  which  ranges 
from  52%  (for  M)  to  22%  (for  N)  of  the  total  sample  variability.  That  signifies 
that  48  to  78%  of  the  total  variability  is  spatially  dependent.  If  we  assume  small 
measurement  errors,  the  large  relative  nugget  for  M  appears  to  be  the  result  of 
very  closely  spaced  variation.  N's  range  is  350  m  and  M's  range  is  525  m, 
indicating  that  our  choice  to  resample  on  a  45-m  grid  was  appropriate.  Results  of 
the  cross  validation  are  summarized  (table  1).  Ordinary  kriging  was  performed  on 
the  grid  points  using  the  16  nearest  neighbors  within  a  radius  of  130  m.  Cross 
validation  statistics  for  median  absolute  error  indicate  that  50%  of  the  obser- 
vations will  be  within  .113  units  (for  M)  and  .111  units  (for  N)  of  the  true  value. 


Table  1. — Stage  2  cross  validation  statistics  for  Moisture  (M)  and  Nutrients  (N). 


Factor 

Mean  Error 

Median  Abs  Error 

MSE 

M 

.001 

.113 

.018 

N 

.000 

.111 

.024 

177 


SITE  SYNECOLOGICAL  COORDINATE  MAPPING 


We  used  the  cross-validated  variogram  models  to  estimate  values  for  M  and  N 
by  applying  ordinary  block  kriging.  Using  the  previous  search  procedure,  we 
produced  contour  maps  (figure  4).  Maps  of  the  synecological  coordinates  were 
compared  with  topographic  relief  in  the  study  area.  No  obvious  relations  with 
elevation  were  found.  The  nutrients  contour  map  shows  the  presence  of  a  large, 
homogeneous  region  to  the  southeast.  Moisture  seems  more  influenced  by  random 
or  micro-relief  variation,  thus  producing  more  irregular  contours  than  nutrients. 


300    400   500  600 
East  (m) 


M  N 
<2.30    I      I  <2.2 

2.30-2.35  f"~1  2.2-2.3 

2.35-2.40 

2.40-2.45 

2.45-2.50 

2.50-2.55 

>2.55 


b) 


300 


400  500 
East  (m) 


600 


Figure  4. — Contour  maps  of  kriged  estimates  of  site  synecological  coordinates  for  stage  2 
data:  (a)  Moisture  and  (b)  Nutrients. 


We  looked  for  similarities  between  the  spatial  distribution  of  each  coordinate 
and  the  inventoried  SI.  The  spatial  patterns  of  kriged  N  appear  to  yield  the  best 
relationship  to  the  three  SI  classes  (figure  5a).  By  plotting  contours  at  2.35  and 
2.55  for  N,  we  delineated  three  areas  with  A  <  2.35  <  B  <  2.55  <  C  (figure  5b), 
which  correspond  roughly  to  areas  with  SI  =  50,  70,  and  80  ft  (figure  5a). 
Estimated  N  coordinates  in  these  upland  aspen  stands  separated  by  .20  units  may 
account  for  SI  changes  of  10  to  20  ft.  Comparing  figure  5a  with  5b  gives  us 
reason  to  support  the  belief  that  vegetative-based  nutrient  estimates  may  be  useful 
in  depicting  areas  of  differing  site  quality. 

DISCUSSION 

Spatial  dependence  is  clearly  evident,  which  allows  for  mapping  site 
synecological  coordinates.  The  N  coordinate  showed  greater  spatial  dependence 
than  the  M  coordinate,  which  showed  a  relative  nugget  effect  of  roughly  50%. 
Viewed  from  another  perspective,  the  variogram  for  M  reveals  two  very  different 


178 


D)  . 

Si 

N 

50 

□ 

A  " 

70 

El 

B 

80 

■ 

c  - 

300     400     500  600 
East  (m) 


300     400     500  600 
East  (m) 


Figure  5 — Comparison  of  site  quality  areas  (a)  as  delineated  by  inventory  site  index 
classes  and  (b)  as  estimated  by  low  (A),  moderate  (B),  and  high  (C)  Nutrient  classes. 


and  yet  equally  strong  aspects  to  the  variability:  50%  is  due  to  measurement  error 
and  very  short-distance  variability  (less  than  5  m),  and  50%  is  due  to  long-range 
spatial  dependence.  For  the  N  coordinate,  the  relative  magnitude  of  the 
variabilitydue  to  the  long-range  spatial  dependence  has  increased  and  the 
magnitude  of  the  micro-site  variability  has  decreased. 

Both  coordinates  showed  ranges  of  at  least  350  m,  which  is  far  greater  than  the 
usual  limits  of  spatial  dependence  in  soil  studies.  Spatial  ranges  for  various  soil 
properties  range  from  1  to  2  m  (Lechowicz  and  Bell  1991,  Jackson  and  Caldwell 
1993)  up  to  15  to  40  m  (Oliver  and  Webster  1991,  Palmer  1990).  The  variation  of 
soils  is  almost  always  more  noisy  or  random  than  other  environmental  variables 
such  as  landform,  climate,  or  geology,  where  long-range  effects  often  dominate 
(Burrough  1983).  The  high  proportion  of  short-range  sources  of  variation  in  soil  is 
caused  by  very  localized  biological  activity,  weathering,  and  small  differences  in 
relief  due  to  glacial  action,  erosion,  and  deposition  (Burrough  1983).  Because  of 
its  large  relative  nugget,  M  appears  more  influenced  by  localized  sources  of 
variation  than  N.  The  large  ranges  and  the  large-scale  patterns  evident  in  figure  4 
suggest  that  broader  trends  are  operating  for  both  factors.  Synecological 
coordinates  have  reduced  the  complexity  of  the  spatial  variation  so  that  micro-site 
variation  in  soil  properties  no  longer  dominates,  but  rather  broader  trends  are 
revealed,  such  as  the  relationship  of  N  coordinates  to  overstory  productivity. 

ACKNOWLEDGMENTS 

Dr.  Cindy  Johnson-Groh,  Janna  Goerdt,  and  Marshall  Kaiser  of  the  College  of 
St  Scholastica  collected  the  plant  lists.  The  St.  Louis  County  Land  Department 
generously  allowed  access  to  its  property,  helped  locate  a  suitable  study  site  (Tom 
Zeisler  and  Mark  Weber),  and  provided  GPS  coordinates  for  the  starting  point 
(Glenn  Berts). 


179 


REFERENCES 


Bakuzis,  E.  V.  1959.  Synecological  coordinates  in  forest  classification  and  in 

reproduction  studies.  Ph.D.  Diss.,  Univ.  of  Minnesota,  St.  Paul.  244  p. 
Bakuzis,  E.  V.,  and  Kurmis,  V.  1978.  Provisional  list  of  synecological  coordinates 

and  selected  ecographs  of  forest  and  other  plant  species  in  Minnesota.  Univ.  of 

Minnesota,  Dept.  of  Forestry.  Staff  Pap.  Ser.  No.  5.  31  p. 
Brand,  G.  J.  1985.  Environmental  indices  for  common  Michigan  trees  and  shrubs. 

USD  A  For.  Serv.  Res.  Pap.  NC-261.  5  p. 
Burgess,  T.  M.,  and  R.  Webster.  1980.  Optimal  interpolation  and  isarithmic 

mapping  of  soil  properties.  II.  Block  kriging.  J.  Soil  Sci.  31:  333-341. 
Burrough,  P.  A.  1983.  Multiscale  sources  of  spatial  variation  in  soil.  I.  The 

application  of  fractal  concepts  to  nested  levels  of  soil  variation.  J.  Soil  Sci.  34: 

577-597. 

Fortin,  M.,  P.  Drapeau,  and  P.  Legendre.  1989.  Spatial  autocorrelation  and 

sampling  design  in  plant  ecology.  Vegetatio  83:  209-222. 
Gutierrez-Espeleta,  E.  E.  1991.  Tropical  forest  site  quality  assessment:  an 

approximation  in  Costa  Rica.  Ph.D.  Diss.,  Iowa  State  Univ.,  Ames.  138  p. 
Jackson,  R.  B.,  and  M.  M.  Caldwell.  1993.  The  scale  of  nutrient  heterogeneity 

around  individual  plants  and  its  quantification  with  geostatistics.  Ecology  74(2): 

612-614. 

Lechowicz,  M.  J.,  and  G.  Bell.  1991.  The  ecology  and  genetics  of  fitness  in  forest 
plants.  II.  Microspatial  heterogeneity  of  the  edaphic  environment.  J.  Ecol.  79: 
687-696. 

Nimerfro,  K.,  and  Brand,  G.  1993.  The  microcomputer  scientific  software  series 

7:  data  recorder  program  for  storing  plant  lists  and  calculating  synecological 

coordinates.  USDA  For.  Serv.  Gen  Tech.  Rep.  NC-157.  20  p. 
Oliver,  M.  A.,  and  R.  Webster.  1986.  Combining  nested  and  linear  sampling  for 

determining  the  scale  and  form  of  spatial  variation  of  regionalized  variables. 

Geo.  Anal.  18:  227-242. 
Oliver,  M.  A.,  and  R.  Webster.  1991.  How  geostatistics  can  help  you.  Soil  Use 

and  Mgmt.  7:  206-217. 
Palmer,  M.  W.  1990.  Spatial  scale  and  patterns  of  species-environment 

relationships  in  hardwood  forest  of  the  North  Carolina  piedmont.  Coenoses  5: 

79-87. 

Webster,  R.,  and  T.  M.  Burgess.  1983.  Spatial  variation  in  soil  and  the  role  of 
kriging.  Agr.  Water  Mgt.  6:  111-122. 

BIOGRAPHICAL  SKETCH 

Margaret  R.  Holdaway  is  a  Mathematical  Statistician,  North  Central  Forest 
Experiment  Station,  St.  Paul,  MN.  She  has  been  a  member  of  the  Station's  Forest 
Inventory  and  Analysis  Unit  since  1995  (previously  with  the  Modeling  Unit)  and 
holds  an  M.S.  in  Biometry  from  the  University  of  Minnesota.  Marge's  areas  of 
interest  are  spatial  statistics,  forest  growth  modeling,  and  climatic  applications. 

Gary  J.  Brand  is  Research  Forester,  North  Central  Forest  Experiment  Station, 
St.  Paul,  MN.  He  has  been  a  member  of  the  Station's  Forest  Inventory  and 
Analysis  Unit  since  1995  (previously  with  the  Modeling  Unit)  and  holds  an  M.F. 
in  Forestry  from  the  University  of  Minnesota.  Gary  is  interested  in  the  relations 
between  synecological  coordinates,  ecological  classifications,  and  forest  stand 
dynamics. 


180 


On  the  Robustness  of  Data 
Assimilation  Methods  in  Air 
Pollution  Modeling 


X.F.  Zhang1  A.W.  Heemink1 


Abstract 

The  performance  robustness  properties  of  two  data  assimilation  meth- 
ods. Kriging  and  Kalman  filtering  presented  for  the  estimation  of  air 
pollution  are  investigated  under  perturbations  of  noise  uncertainty  and 
model  parameter  uncertainty.  The  simulation  results  are  used  to  illus- 
trate the  advantages  of  Kalman  filtering  approach  under  various  cir- 
cumstances. 


INTRODUCTION 

In  Zhang  et.  al  (1995)  two  data  assimilation  methods,  a  Kriging  approach 
and  Kalman  filtering  are  developed  for  tbe  estimation  of  two  dimensional  air 
pollution  problem.  The  deterministic  system  model  and  the  statistics  of  system 
noise  and  measurement  noise  are  assumed  to  be  known  exactly  in  Zhang  et. 
al  i  199*)).  However,  such  an  assumption  is  often  unrealistic  in  applications, 
and  noticeable  deterioration  of  system  performance  may  be  caused  by  model 
errors  and  noise  uncertainty. 

In  this  paper  we  investigate  the  performance  robustness  properties  of  the 
two  data  assimilation  methods  proposed  in  Zhang  et.  al  (1995).  The  effects 
of  both  noise  uncertainty  and  parameter  uncertainty  are  examined.  In  Section 
2,  the  stochastic  system  model  is  described,  and  in  Section  3  the  two  data 
assimilation  methods  are  briefly  reviewed.  In  Section  4,  the  influence  of  noise 
uncertainty  on  the  Kalman  filtering  approach  is  analysed,  and  a  strategy  to 
minimize  the  worst  case  performance  is  proposed.  Then,  in  Section  5,  sim- 
ulation results  are  presented,  showing  the  advantages  of  the  proposed  robust 
Kalman  filter  design  method  compared  with  the  conventional  Kalman  filter 
design  and  the  Kriging  approach  at  present  of  noise  uncertainty.  Furthermore, 
the  performance  robustness  is  checked  with  respect  to  uncertain  transport 
speed  of  air  pollution.  Finally,  Conclusion  is  given  in  Section  6. 


1  Department  of  Technical  Mathematics  and  Informatics,  Delft  University  of  Technology, 
Mekelweg  Jf.  P.O.  Box  5031,  2600  GA  Delft,  The  Netherlands 


181 


THE  STOCHASTIC  DYNAMIC  AIR  POLLUTION  MODEL 

The  transport  phenomena  of  air  pollutant  in  the  environment  are  usually 
described  by  an  advection-diffusion  equation 

^  =  -\7-[uc-\7-Dc\  (1) 

where  c  is  the  concentration  of  the  pollutant,  t  is  time,  u  is  the  velocity  field, 
D  is  the  dispersion  tensor  including  molecular  diffusion  and  dispersion.  By 
introducing  a  numerical  scheme  (Heemink  1990,  Van  Eijkeren  1993)  Eq.  (1) 
can  be  represented  by  a  discrete  time  state  space  model. 

x(k)  =  $x(k-l)  +  rw(k)  k  =  1,2, ...  (2) 
z(k)   =   Hx{k)  +  v{k)  (3) 

where  x(k)  is  an  n- vector  state  process  taking  corresponding  values  at  model 
grid  point  at  time  instant  A:,  n  is  the  total  number  of  model  grid  points  at  the 
pollutant  concentration  field,  and  $  is  an  n  x  n  dimension  transient  matrix 
obtained  by  the  numerical  scheme  discretizing  the  advection-diffusion  Eq.  (1), 
T  is  an  n  x  q  noise  input  matrix  that  interpolates  q  dimensional  system  noise 
w(k)  at  n  model  grid  points,  z(k)  is  an  m-vector  measurement  process,  with 
measurement  noise  sequence  v(k).  <I>,  T  and  H  are  constant  matrices,  w(k) 
and  v(k)  and  the  initial  state  x(0)  are  assumed  to  be  multiply  independent 
and  satisfied  the  following  conditions: 

E{w(k)w'(k)}  =  QSjk,  E{v(k)v'{k)}  =  R6,k  (4) 
E{x{Q)}  =  S(0),  E{1*(0)  -  S(0)j[a>(0)  -  4(0)]'}  =  P0  (5) 

where  Q  >  0  is  a  positive  semi  definite  constant  matrix,  R  >  0  is  a  positive 
definite  constant  matrix,  Sjk  =  1  if  j  =  k  and  Sjk  =  0  if  j  k. 

DATA  ASSIMILATION  METHODS 

m 

The  Kriging  Approach 

Kriging  is  a  method  for  the  optimal  estimation  of  a  quantity  y  at  an  ar- 
bitrary location  that  has  not  been  measured  by  using  measurements  taken  at 
surrounding  points.  By  changing  the  position  of  the  point  of  estimation,  it  is 
possible  to  estimate  the  whole  field  of  spatial  variables,  the  estimation  of  the 
pollutants  concentration  value  x  at  grid  point  of  Gm  and  time  instant  k  is 
given  by: 

*t-(jfe|jb)    =  $(i,:)x(k-l\k-l) 

+  E ~      0«(*  1)] 

i  =  1, 2, n  (6) 


182 


where    is  t  he  total  number  of  measurement  stations.  To  determine  the  optimal 

choices  of  K(i,  1),  A'(/.  2)  /\'(/,  .s)  of  minimizing  the  estimation  variance  for 

an  arbitrary  ith  grid  point  of  Gm-  We  get  (Zhang  et.  al  1995) 

J2(l'(hji)- RJi)I<(iJ)  +  fi=l(hJl)   j  = 
/=i 

(7) 

i=l 

where  7^/  denotes  the  value  of  the  semi-variogram  'y(hjj)  of  the  deviation  field. 
Rji  denotes  the  value  of  the  known  covariance  of  measurements  noise.  The 
spatial  correlation  in  the  form  of  a  semi-variogram  is  automatically  estimated 
based  on  the  available  measurement  data  at  each  time  instant.  The  set  of 
K(i,j),  j  =  l....,.s  is  obtained  by  Eq.  (7)  and  then  applied  to  Eq.  (6).  This 
produces  an  unbiased  estimation  with  minimum  variance  of  estimation  error. 
The  updated  picture  of  the  whole  concentration  field  is  obtained  by  adjusting 
xt  throughout  every  grid  point  of  the  concentration  field  using  Eq.  (6). 

The  Kalman  Filter  Approach 

Since  the  linear  discrete-time  system  Eq.  (2)  and  Eq.  (3)  is  time-invariant, 
the  Chandrasekhar-type  algorithm  (Morf  et  al  1974)  can  be  employed  to  get  a 
steady  state  Kalman  filter  as  follows: 


x(k\k)  = 

=  $x(k\k  -  1)  +  K[z(k)  -  Hx(k\k  -  1)] 

(8) 

x(k\k  — 

1)  =  $x(k  -  \\k  -  1) 

(9) 

where  the  gain  matrix  can  be  obtained  by  (Zhang  et.  al  1995): 

G(k+\)  = 

G(k)  +  Y(k)L(k)Y(k)'H' 

(10) 

R?(k  +  l)  = 

Re(k)  +  HY(k)L(k)Y(k)'H' 

(11) 

K{k  +  l)  = 

G{k  +  \)RE(k+\)-1 

(12) 

+  = 

Y{k)  -  K(k+\)HY(k) 

(13) 

L(k  + 1)  = 

L{k)  +  L(k)Y(k)'H'Re(kylHY{k)L(k) 

(14) 

Y(k  +  1)  = 

*S(k  +  l) 

(15) 

with  initial  condition: 

Y(\)  =  T,G(Q)  =  0,Re(0)  =  R,L{0)  =  Q  (16) 

The  equations  from  Eq.  (10)  to  Eq.  (15)  are  iterated  until: 

\\K(k+\)-K(k)\\<9\\K(k)\\  (17) 

where  6  is  determined  by  prespecified  accuracy.  Under  the  assumption  that  the 
dimension  of  system  noise  process  q  is  much  less  than  the  dimension  of  system 
states  n.  i.e,  q  <C  7i,  the  Chandrasekhar-type  algorithm  provides  a  significant 
computation  reduction  (Heemink  1988,1990). 


183 


A  ROBUST  FILTER  DESIGN  METHOD  UNDER 
PERTURBATIONS  OF  NOISE  UNCERTAINTY 


In  general,  it  is  a  hard  task  to  get  suitable  Q  and  R  matrices  which  fit  best 
the  stochastic  system  model  Eq.  (2)  with  various  observations.  We  suppose 
that  the  real  steady  state  covariances  of  system  noise  and  measurement  noise 
are  uncertain,  and  they  can  be  described  by:  Q  =  Qo  -f  AQ,  R  =  R0  +  AR. 
Further,  we  suppose  that  all  information  on  AQ  and  AR  is  that  the  norm  of 
AQ  and  AR  are  bounded  by:  ||AQ||  <  a,  \\AR\\  <  (3.  As  the  optimal  gain 
matrix  K  of  the  Kalman  filter  is  a  function  of  Q,  and  R  ,  the  K  obtained 
by  using  Q0  and  R0  is  in  general  no  longer  optimal  due  to  the  existence  of 
uncertainties  in  Q  and  R  matrices.  First,  we  consider  the  prediction-type 
Kalman  filter  with  gain  matrix  Kv.  Let  e(k  +  1)  =  x(k)  —  x(k\k  —  1)  the 
actual  estimation  covariance  denoted  by  N°(k  +  l\k)  is  determined  by 

N°(k  +  l\k)    =   E{e{k  +  l\k)e'{k+  l\k)} 

=   ($  -  KpH)N°(k\k  —  1)($  -  KPH)' 

+   r(Q0  +  AQ)r'  +  Kp(Ro  +  AR)K'p  (18) 

Sangsuk-Iam  (1990)  prove  that  as  k  — >  oo  Eq.  (18)  always  converges  to  some 
constant  TV0  which  satisfies  the  following  algebraic  Riccati  equation 

N°  =  ($  -  KPH)N°($  -  KpH)'  +  T(Q0  +  AQ)F  +  KP(R0  +  AR)K'p  (19) 

From  Eq.  (19)  we  notice  that  the  actual  estimation  covariance  TV0  is  a  mono- 
tonic  increasing  function  of  the  actual  noise  covariances  (Jo  +  AQ  and  Rq-\-AR. 
With  Q  :=  AQ0  +  al,  R  :=  ARo  +  (51  where  /  is  an  identity  matrix,  we  have 
Q  ^  Qo  +  AQ  >  Ro  +  AR.  It  is  a  reasonable  choice  to  take  the  worst  case 
covariance  Q  and  R  to  design  the  Kalman  filter  which  minimizes  the  upper 
bound  of  Kalman  filter  performance,  i.e 

N{K;,Q,R)  <  N(KP,Q,R)  (20) 

where  K*  is  the  optimal  gain  matrix  of  the  Kalman  filter  designed  for  the  fixed 
pair  of  worst  case  noise  covariances  Qo  +  and  Ro  +  (31 }  and  Kv  could  be 
any  other  gain  matrix.  We  have  the  following  saddle  point  inequality: 

n(k;,q0  +  aq,r*  +  ar)  <  n(k;,q,r)  <  n(kp,q,r)  (21) 

Now,  we  consider  the  current  estimation-type  Kalman  filter  with  gain  matrix 
Ke.  From  Eq.  (2),  Eq.  (3),andEq.  (8)  we  have  the  estimation  error  covariance 
of  the  form: 

P(k)   =   (I  —  KeH)$P(k  —  1)<J>'(7  -  KeH)' 

+   (/  -  KeH)TQY\I  -  KeH)'  +  KeRK'e  (22) 


184 


Obviously,  for  any  determined  7\e,  the  actual  estimation  covariance  P(k)  is  also 
a  monotonic  increasing  function  of  Q  4-  AQ,  and  R  +  A/?.  We  have  following 
saddle  point  inequality 

P(K*j  Q0  +  AQ,  /?0  +  A/?.)  <  P(I<;,Q,R)<N(Ke,Q.~R)  (23) 

where  A'*  is  the  optimal  gain  matrix  of  the  current  estimation-type  Kalman 
filter  designed  using  the  fixed  pair  of  worst  case  noise  covariances  Q0  +  a  /  and 
Ro  +  (II ,  and  7\"f  could  be  any  other  gain  matrix.  From  above  discussion  we  can 
see  that  the  selection  of  worst  case  noise  covariances  is  a  strategy  to  design  a 
robust  Kalman  filter  (either  prediction-type  or  current  estimation-type)  which 
minimizes  the  upper  bound  of  the  Kalman  filter  performance. 

SIMULATION  STUDY 

The  simulation  studies  are  aimed  to  examine  the  performance  robustness  of 
the  two  data  assimilation  methods  with  respect  to  noise  uncertainty  and  model 
parameter  uncertainty.  The  experimental  region  is  a  cycle  area  bounded  in  ft 
(see  figure  1 ). 

(a):  True  field  (b):  9  Sample  Locations 


south  south 


Figure  I:  (a):  The  concentration  field  at  t=64;  (b):  "*"  Observations,  "." 
Model  grid  Gm;  (c),  (d):  The  corespponding  results  of  the  Filtering  and 
the  Kriging  approaches  without  model  parameter  and  noise  uncentainty. 


185 


The  concentration  of  the  pollutant  is  denoted  by  x(£,t)  where  £  is  the  co- 
ordinate £  =  (x,  y)'  G  H.  The  true  field  is  a  stochastic  process  described  by  Eq. 
(2).  The  deterministic  part  of  this  stochastic  process  is  driven  by  the  advection- 
diffusion  equation  Eq  (1)  with  a  circular  velocity  it  =  u((x,  y)')  =  27r(— y,^)'  by 
using  the  numerical  approximation  described  in  Heemink  (1990)  and  Van  Eijk- 
eren  (1993).  The  initial  condition  of  Eq.  (1)  consists  of  a  constant  background 
field  and  a  small  cosine-squared  shape  cone  interpreted  as  a  high  density  con- 
centration area  of  local  pollution.  In  our  experiments,  the  simulated  true  field 
at  t  =  64  is  shown  in  figure  1,  (a).  The  regular  model  grid  Gm  is  shown  in 
figure  1,  (b).  The  system  noise  w  is  generated  by  w  =  (w-[ , wq)'  with  zero 

1     17 x  _x  \2  I  /     _  \2l 

mean  and  constant  covariance  Qo(i,  j)  =  2  x  10    e  0  06  u-'  -J'    v-«  V    and  the 

measurement  noise  v  is  generated  by  Ro(i,j)  =  1  x  10-2e  001  ~^  K 
The  location  of  9  observation  stations  is  also  plotted  in  figure  1  (b).  When 
the  deterministic  model  ,  the  statistics  of  system  noise  and  measurement  noise 
are  exactly  known  the  reasonable  good  results  of  both  data  assimilation  ap- 
proaches are  plotted  in  figure  1  (c),  (d). 

Experiment  1 

We  assume  that  the  second  moments  of  statistics  of  the  system  noise  and  mea- 
surement noise  are  uncertain.  Let  nominal  filter  use  the  optimal  gain  matrix 
corresponding  to  the  nominal  system  noise  covariance  Qo  and  measurement 
noise  covariance  Ro,  and  robust  filter  use  the  robust  gain  matrix  which  min- 
imize the  worst  case  of  Q  =  Qo  +  al  and  R  =  Ro  +  /3I  with  a  =  1  x  10~4 
and  (3  =  5  x  10-5.  First  Q  =  Q0  and  R  =  R0:  The  Root  Mean  Square  (RMS) 
of  the  nominal  filter  and  robust  filter  are  plotted  in  (a)  of  figure  2.  It  can  be 
seen  that  the  RMS  corresponding  to  robust  filter  is  only  slight  higher  than 
that  of  nominal  filter  which  is  the  optimal  gain  matrix  in  this  case.  Secondly 
Q  =  Qo  +  olI  and  R  =  R0  +  (51:  The  RMS  of  the  nominal  filter  and  robust 
filter  are  plotted  in  (b)  of  figure  2.  We  can  see  that  the  RMS  of  robust  filter 
is  now  much  lower  than  that  of  nominal  filter.  It  is  obvious  from  above  sim- 
ulation results  the  robust  filter  which  minimize  the  worst  case  covariance  of 
the  estimation  error  makes  the  performance  of  the  Kalman  filter  much  better 
from  the  point  of  view  of  robustness.  To  compare  the  two  data  assimilation 
methods  the  corresponding  RMS  of  the  Kriging  approach  are  also  plotted  in 
figure  2  (a),  (b).  It  can  be  seen  that  Kalman  filter  approach  can  provide  more 
accurate  result  then  The  Kriging  approach. 

Experiment  2 

We  suppose  that  the  Kalman  filter  is  perturbed  by  uncertain  deterministic 
system  model.  More  exactly,  it  is  assumed  that  the  transport  speed  u  which 
is  a  key  model  parameter  is  uncertain.  Let  the  Kalman  filter  be  designed  by 
^o  =  {ux,uy)  =  27r(— y,x),  and  let  the  true  concentration  field  is  run  with 


186 


u  =  uQ  +  Aw.  Aw  =  — O.ltto  i.e.  10%  deviation.  The  true  concentration  field 
at  /  =  ()4  is  plotted  in  (c)  of  figure  2.  The  corresponding  estimated  field  of  the 
Kalman  filter  approach  and  the  Kriging  approach  are  plotted  in  (d)  and  (e)  of 
figure  2,  respectively.  The  RMS  of  the  data  assimilation  procedures  are  plotted 
in  (f)  of  figure  2.  It  is  clear  in  this  case  that  the  performance  robustness  of 
the  Kalman  filter  approach  to  tolerate  model  parameter  uncertainty  is  much 
better  than  that  of  the  Kriging  approach. 


(a):  RMS  (b):  RMS 


Figure  2:  (a)  and  (b):  The  RMS  of  three  methods  with  noise  uncertainty, 
(c):  Truth  field  at  t=64;  (d)  and  (e):  The  corresponding  results  of  Fil- 
tering and  Kriging  approaches  with  10%  model  parameter  uncertainty; 
(f ):  RMS  of  both  two  approaches. 


187 


CONCLUSION 


In  this  paper,  the  performance  robustness  of  two  data  assimilation  methods 
under  the  perturbations  of  noise  uncertainty  and  parameter  uncertainty  is  in- 
vestigated in  two  dimensional  air  pollution  problem.  The  experimental  results 
show  the  advantage  of  a  robust  Kalman  filter  design  method  with  respect  to 
noise  uncertainty.  Moreover,  it  is  demonstrated  in  the  simulation  studies  that 
the  Kalman  filter  approach  has  much  stronger  performance  robustness  to  tol- 
erate model  parameter  uncertainty  then  the  Kriging  approach.  Our  current 
research  focuses  on  applying  Kalman  filtering  to  "  the  evaluation  of  CH4  bud- 
get in  Europe  "  in  cooperation  with  RIVM  air  quality  laboratory. 

ACKNOWLEDGMENTS 

This  work  has  been  carried  out  in  cooperation  with  and  with  financial  support 
from  the  RIVM. 

REFERENCES 

Heemink,  A.W.,  1988.  Two-Dimensional  Shallow  Water  Flow  Identification. 

Applied  Mat.  Mod.,  109-118. 
Heemink,  A.W.,  1990.  Identification  of  Wind  Stress  on  Shallow  Water  Flow 

Surfaces  by  Optimal  Smoothing.    Stochastic  Hydrol.    Hydraul.   Vol.  4, 

105-119. 

Morf,  M.,  Sidhu,  S.S.,  Kailath,  T.,  1974.  Some  New  Algorithms  for  Recursive 
Estimation  in  Constant,  Linear,  Discrete-Time  system.  Trans.  Autom. 
Control,  Vol.  AC-19,  NO.  4,  315-323. 

Sangsuk-Iam,  S.,  1990.  Analysis  of  discrete-time  Kalman  filtering  under  incor- 
rect noise  covariance,  IEEE  Transactions  on  Automatic  Control,  Vol.  35, 
1304-1309. 

Van  Eijkeren,  J.C.H.,  1993.  Backward  Semi-Lagrangian  Methods:an  Adjoint 
Equation  Method.  In:  Numerical  Methods  for  Advection  -  Diffusion  Prob- 
lems, Vreugde  hil  OB  and  Kora  B  Eds.  Vieweg,  Vol,  45,  215-241. 

Zhang,  X.F.,  Van  Eijkeren,  J. OH.  and  Heemink,  A.W.  1995  Data  Assimilation 
in  Dynamic  Environmental  Pollution  Modeling.  RIVM  report,  No.  421503 
006,  Bilthoven,  The  Netherlands. 

BIOGRAPHICAL  SKETCH 

Zhang  X.F.  is  doing  a  Ph.D.  project  on  the  development  of  data  assimilation 
methods  for  the  estimation  of  air  pollution. 

Heemink  A.W.  is  a  professor  of  faculty  of  technical  mathematics  and  infor- 
matics, TU  Delft. 


188 


The  Development  of  a  GIS 
Information  Quality  Module 

Jane  Drummond    Allan  Brown  2,  Du  Daosheng  3,  Corne  van 

Elzakker  4 


Abstract.-  In  the  early  1980s  the  authors  first  considered  1)  the 
transfer  of  traditional  mapmakers'  expertise  regarding  spatial 
information  quality  into  GIS  and  2)  ways  of  determining  GIS 
generated  information's  quality.  Taking  into  account  the  now  accepted 
principles  that  attribute,  position,  age,  and  topology  quality  and 
lineage  all  contribute  to  GIS  generated  information,  the  framework 
for  our  work  was  that  as  a  GIS  generated  new  information  an 
associated  uncertainty  subsystem  should  in  parallel  generate 
information  on  the  quality  of  that  information  and  that  the  two  sets  of 
information  should  be  displayed  simultaneously  to  better  the  GIS 
user's  decision  making.  Now  an  extant  (though  incomplete)  entity  the 
subsystem  consists  of  appropriate  means  to  determine  and  store  the 
quality  of  spatial  data  and  the  processing  model  used  to  generate 
information,  means  for  determining  the  generated  information  quality 
(variance  propagation,  fuzzy  sub-set  theory,  etc.)  and  means  for 
visualizing  that  quality.  These  are  integrated  in  an  Uncertainty 
Subsystem  called  PUIUS.  The  subsystem  has  been  implemented  in 
ILWIS.  This  paper  will  review  the  subsystem  (and  another), 
summarize  some  applications  and  point  the  way  to  further  work. 


INTRODUCTION 

Maintaining  high  standards  has  long  concerned  land  surveyors, 
photogrammetrists  and  cartographers,  but  the  Geographic  Information  System 
(GIS)  community  seemed  only  alerted  to  the  problem  of  information  quality  in 
the  early  1980's.  By  the  end  of  the  1980's  those  in  GIS  had  accepted  the 
principles  that:  quality  of  attribute  and  position;  information  age;  topological 
consistency;  and  lineage  all  contributed  to  the  quality  of  any  GIS  generated 
information.  We  first  started,  formally,  to  consider  1)  the  transfer  of 
traditional  mapmakers'  expertise  regarding  spatial  information  quality  in  the 

'  Jane  I )rummdnd,  Topographic  Science  Section.  Glasgow  I  niversity:  I  K 
i 

'  Allan  Hrown.  Department  oj 'Geoinformatics,  ITC.  Enschede,  The  Netherlands 

~y  Du  Daosheng.  Professor -Wuhan  Technical  University  of  Surveying  and  Mapping,  Wuhan,  PR  China 

~*  t  orne  van  Elzakker.  Department  of  Geoinformatics.  TTT,  Enschede,  The  Netherlands 


189 


GIS  environment  and  2)  ways  to  determine  the  quality  of  GIS  generated 
information,  in  the  mid-eighties. 

Taking  into  account  the  principles  presented  in  the  first  paragraph,  our 
study's  framework  was  that  while  a  GIS  generated  new  information  an 
associated  uncertainty  subsystem  should  in  parallel  generate  information  on 
the  quality  of  that  new  information  and  that  the  two  sets  of  information  should 
(if  possible)  be  displayed  simultaneously  to  better  the  GIS  user's  decision 
making.  The  subsystem  itself  would  consist  of  appropriate  means  to  determine 
(knowledge  base,  user  queries,  etc.)  and  store  the  quality  of  spatial  data,  to 
store  the  quality  of  the  processing  model  used  to  generate  information,  the 
means  for  determining  the  generated  information  quality  (variance 
propagation,  fuzzy  sub-set  theory,  etc.)  and  the  means  (2D,  3D)  for  visualizing 
that  quality.  These  were  integrated  in  an  Uncertainty  Subsystem  called  PUIUS 
(Prototype  User  Interface  for  an  Uncertainty  Subsystem).  The  subsystem  has 
been  implemented  in  ILWIS,  a  PC  based  GIS  developed  and  marketed  bt  the 
International  Institute  for  Aerospace  Survey  and  Earth  Sciences  (ITC)  in  The 
Netherlands.  This  paper  will  review  this  subsystem. 

The  review  will  take  the  form  of  a  summary  of  projects  carried  out  by  the 
group  which  lead  to  the  development  of  PUIUS,  and  an  investigation  of  a 
recent  version  of  another  off-the-shelf  GIS  package  which  is  now  marketted 
with  some  information  quality  processing  tools.  Thus  the  five  remaining 
sections  of  this  paper  will  deal  with  i)  early  ideas;  ii)  the  XGIS  project;  iii)  the 
PUIUS  project;  iv)  the  Idrisi's  information  quality  processing  tools;  v)  and 
conclusions. 

EARLY  IDEAS 

In  1983  two  of  this  paper's  authors,  working  at  the  UK  Experimental 
Cartography  Unit,  prepared  a  very  straightforward  presentation  on  digital 
spatial  data  quality  for  a  Visiting  Committee.  The  presentation  showed 
boundary  widths  for  polygons  of  various  natural  features  as  a  function  of  the 
standard  deviation  (SD)  of  these  boundaries'  coordinates,  and  then  the 
resulting  overlaid  polygons.  Some  visitors  expressed  surprise  at  the 
demonstrated  imprecise  nature  of  the  digital  spatial  data  we  so  expensively 
acquired.  Although  official  British  science  seemed  unready  pay  for 
recognisably  imperfect  digital  spatial  data  (the  two  authors  concerned 
relocated  to  the  Netherlands  and  China),  at  about  the  same  time  Blakemore 
(1983)  and  Chrisman  (1983)  were  much  more  successfully  convincing 
European  and  North  American  academics  that  the  imperfections  of  digital 
spatial  data  represented  a  respectable  research  area.  The  presentation's 
polygons  represented  areas  of  (assumed)  uniform  slope,  soil  class,  rainfall  and 
temperature  and  were  overlaid  and  used  for  crop  suitability  mapping 


190 


(ODA.1983).  At  the  time,  and  working  in  the  vector  environment,  to 
determine  the  precision  of  such  contributing  polygons'  boundaries  and  then 
produce  overlaid  polygons  with  boundary  line  widths  proportional  to  their 
(e.g.)  precision  was  cumbersome.  The  same  problem  has  now  been  much 
more  elegantly  addressed  in  the  raster  environment  of  PUIUS  (see  below). 

The  user  of  a  GIS  is  not  likely  to  be  interested  in  the  quality  of  the  data  per 
se,  but  in  the  quality  of  the  generated  information.  In  the  ODA  suitability 
mapping  project,  the  input  sets  of  polygons  each  had  a  centroid  -  as  had  each 
polygon  resulting  from  overlay.  Standard  databasing  facilities  permit  such 
centroids  to  have  attributes  referring  to  data  quality  and  can  be  easily 
processed.  Such  was  part  of  the  thinking  behind  much  subsequent  work  in  the 
group  (Drummond,  1987).  The  group  understood  a  GIS  to  be  both: 

(i)  a  system  which  used  processing  models  to  generate  new 
information  from  existing  data;  and, 

(ii)  a  spatial  data  archive  which  allowed  one  to  retrieve  and  display 
stored  data. 

In  the  first  case  the  new  information  will  be  visualized,  but  simultaneously 
(or  by  toggling)  its  quality  should  be  visualised.  Error  propagation  models  are 
directly  related  to  information  generation  models  and  information  generation 
models  can  be  mathematical  (e.g.  computation  of  aspect;  USLE)  or  logical 
(e.g.  crop  suitabilty;  Community  Tax  determination).  Thus  if  a  mathematical 
information  processing  model  gives  us  X  =  f(a,b,c)  then  oX  =  f(aa,ab,ac), 
where  a  represents  Standard  Deviation.  Likewise  if  a  logical  processing  model 
gives  us  Y  =  f  (l,m,n)  then  pY  =  f(pl,pm.pn)  where  p  represents  a  certainty 
statistic  such  as  Probability  or  Certainty  Factor.  However  the  processing 
model  itself  may  be  in  error,  thus  however  good  the  data  are,  the  generated 
information  may  be  poor.  In  the  case  of  a  mathematical  model  its  error  can  be 
handled  as  noise  and  subsequently  processed  by  oX  =  f(oa,ab,oc,on)  where  n 
represents  noise.  In  the  case  of  a  logical  model,  the  model  itself  may  have  a 
certainty  statistic  associated  with  it  which  can  be  processed  by  pY  = 
f(  pl.pm.pn.po)  with  po  representing  probability  or  CF  of  the  model  holding. 

In  the  second  case  above  a  GIS  might,  e.g.,  be  queried  to  produce  a  map  to 
show  the  position  of  a  utility  line  through  a  neighbourhood.  The  users  of  the 
map  might  be  concerned  with  the  accuracy  of  depiction  of  the  utility's 
location.  If  the  precision  of  the  procedure  used  to  survey  in  the  utility  was 
recorded  in  the  attribute  tables  of  the  GIS,  then  this  information  could  be 
extracted  and  used  to  control  the  width  (e.g.  3a  allowing  >99.9%  probability 
of  finding  the  utility)  of  the  utility's  mapping.  In  this  case  no  processing 
model  is  used  to  generate  new  information. 

Following  our  understanding  of  GIS,  a  simple  GIS  using  the  merged 
capabilities  of  AutoCAD  and  dBASE  (Drummond,  1990)  was  built.  As  well  as 
more  usual  facilities,  this  GIS  had  an  interface  permitting  interrogation  of  the 


191 


user  to  acquire  quality  information  of  each  data  set.  Although  not  a  very  user- 
friendly  interface  (when  compared  to  PUIUS  described  in  a  later  section)  it 
permitted  the  user  to  provide  the  quality  information  of  each  data  item  in  a, 
probablity  or  natural  language  terms  convertible  to  CF,  as  appropriate. 
Conversion  (for  example  converting  a  to  probability  using  central  limit  theory 
when  a  contiuous  variable,  such  as  rainfall,  was  reclassified  into  a 
discontinuous  variable  such  as  the  rainfall  class  300-400  mm  a  year)  was 
carried  out  within  the  GIS.  The  acquired  information  was  stored  in  the 
relevant  attribute  database  table  and  processed  using  fuzzy  sub-set  reasoning 
to  provide  CF  values  for  each  crop  suitability  polygon  (Drummond,1987).  As 
the  data  concerned  (i.e.  for  the  crop  suitability  project)  were  all  of  a 
discontinuous  type  error  propagation  using  fuzzy  logic  rather  than  variance 
propagation  was  appropriate  -  but  variance  propagation  was  supported.  Simple 
maps  showing  the  quality  of  generated  suitability  information,  by  polygon  and 
with  quality  a  function  of  hatching  density,  were  produced. 


FIGURE  1 

UNCERTAINTY  SUBSYSTEM. 
The  overall  concept. 


FIGURE  2 

UNCERTAINTY  SUBSYSTEM 
Visualization  of  Information 
Quality 


Uncertainty  Sub-System 

Top-level 


Processing 
Model 


Storing/Determining 
quality  of  Data 

A 


Storing 
quality  of  the 
model  q 


Determining  of 
Final  information 
Quality  r 


Visualization  of 
Quality 
A    Information  q 


fQuality  mao 


fQuality  repon 


Uncertainty  Sub-System  (D) 

Level  1 :  Visualization  of  Quaiitv  information 


~mai  iniormaoon  duality  rile 


Cartograonic 


Non- 

Cartograonic 


Reoresent  trie: 
point  acc.-> 

Point  Size 
line  acc.  •> 

Line  Width 


T 

T 

Choose 

Insert  the 

Visual  vanaoie: 

different 

-  Texture 

Quality 

•  Size 

Indicators 

•Value 

in  a  reoort 

Dtsoiay  the 
Quality  info, 
in  a  3ar-cnart 


Oisolay  the 
Quality  info. 
Mao  (2D  or  3D) 


''Quality 

'  Quality 

'Quality 

Mao 

Bar-Chart 

Reoort 

.qua 


192 


THE  XGIS  PROJECT 


The  ITC  was  willing  to  partially  support  three  MSc  students 
(Ramlal,1991 ),  (Fan,  1992),  (01iveira,1992)  to  research  data  quality  matters  in 
their  'dissertation  year'  and  fund  a  research  associate  (Barsoum,1995).  This 
was  referred  to  as  the  XGIS  project  and  represented  the  development  of  an 
Uncertainty  Subsystem  -  a  subsystem  of  ILWIS  (Ramlal  et  al.,  1991).  The 
overall  concept  of  this  subsystem  is  provided  in  FIGURE  1. 

From  FIGURE  1  it  can  be  seen  that  the  Uncertainty  Subsystem  was 
designed  with  four  components.  These  dealt  with  the  uncertainty  of  data 
(Storing  Determining  quality  of  data),  the  uncertainty  of  processing  models 
(Storing  quality  of  the  model),  error  propagation  (Determining  of  Final 
Information  Quality)  and  the  visualisation  of  uncertainty  (Visualisation  of 
Information  Quality).  The  visualisation  was  tackled  first  (Ramlal,  1 99 1 ), 
(Ramlal  et  al.,1992),  (Van  Elzakker  et  al.  1992).  FIGURE  2  summarises  this. 

From  FIGURE  2  it  can  be  seen  that  the  visual  variables  texture,  size  and 
value  were  considered  appropriate  for  depicting  data  quality.  The  above 
mentioned  work  by  Ramlal,  Van  Elzakker  and  Drummond  resulted  in  maps  of 
a  type  in  which  suitability  was  represented  by  the  visual  variable  colour  (e.g. 
green  =  most  suitable  to  red  =  least  suitable)  and  data  quality  by  value  (e.g. 
dark  most  certain  to  light  least  certain).  Later  Brown  and  Van  Elzakker 
(Brown  et  al..,  1 993)  experimented  with  the  use  of  colour  in  the  representation 
of  categoric  area  information  quality.  They  concluded  that  the  simultaneous 
representation  of  attribute  and  quality  is  possible  (by  means  of  the  colour  hue 
and  colour  saturation  respectively),  for  both  screen  display  and  printer 
output,  for  a  limited  number  of  categories  and  quality  levels. 

In  tackling  model  quality  (Fan,  1992),  Fan  took  on  the  task  of  developing 
Part  B  (FIGURE  1 )  of  the  Uncertainty  Sub-System  addressing  (regression) 
model  building  within  the  GIS  environment.  By  using  Linear  Least  Squares 
Regression  in  model  building,  information  became  available  on  the 
performance  (quality)  of  the  model.  This  information  was  graphically 
displayed  allowing  the  model  developer  to  improve  the  model.  Actual  model 
development  related  to  toxic  cloud  release  was  carried  out.  Fan  established  the 
concept  of  a  model  library  in  ILWIS  where  developed  and  pre-existing  models 
could  be  stored  along  with  relevant  quality  and  other  information. 

De  Oliveira  addressed  data  quality  (01iveira,1992)  (Storing/ Determining 
Data  Quality  in  FIGURE  1 )  and  produced  an  appropriate  module  for  the  sub- 
system. The  module  is  capable  of  automatically  generating  data  quality 
parameters  from  related  'Ground  Truth'  data  and  user-supplied  information. 
The  algorithms,  in  the  main,  concerned  positional  and  attribute  accuracy 
assessment,  working  at  two  levels.  The  first  level  referred  to  overall  accuracy 
parameters  at  the  data  set  level  (e.g  theme)  and  the  second  represented 
accuracy  parameters  associated  with  individual  database  objects.  It  was  a 


193 


requirement  of  De  Oliveira's  future  employer  (University  of  Aveiro)  that 
concepts  developed  and  implemented  for  ILWIS  could  also  be  implemented  in 
Arc/Info  to  support  an  Emergency  Response  GIS.  Error  propagation  for  the 
particualr  information  generation  (or  processing)  models  used  (relating  to 
chemical  accidents  in  the  form  of  toxic  cloud  releases)  was  developed. 

The  challenge  of  developing  a  complete  error  propagation  module  (i.e.  Part 
C  of  the  Uncertainty  Sub-System:  Determination  of  Final  Information  Quality 
-  see  FIGURE  1)  has  remained  unaddressed  in  the  XGIS  project.  Instead  the 
remaining  funds  were  used  to  support  development  of  the  user  interface 
(within  the  PUIUS  Project).  User  interface  is  an  important  issue  if  GIS  users 
are  to  give  serious  consideration  to  matters  of  data  and  information  quality. 

THE  PUIUS  PROJECT 

In  its  original  form  ILWIS  operated  in  the  MS-DOS  environment.  With  the 
decision  to  market  ILWIS  for  WINDOWS  from  1995  it  was  necessary  to 
transfer  XGIS  into  WINDOWS.  In  its  prototype  form  and  in  the  WINDOWS 
environment  the  ILWIS  Uncertainty  Subsystem  has  the  name  of  PUIUS 
(Prototype  Uncertainty  Subsystem).  It  was  hoped  that  the  PUIUS  project 
would  complete  the  XGIS  development,  at  least  in  prototype  form.  Referring 
to  FIGURE  1  the  transfer  of  parts  B  and  D  was  completed,  and  of  part  C  for 
discontinuous  variables,  but  not  continuous  variables  (Barsoum,1995). 

The  prototype  was  developed  using  Visual  Basic.  The  prototype  deals  with 
the  assessment  and  representation  of  the  quality  of  discontinuous  (categoric) 
attribute  data  and  resulting  information  referring  to  polygons.  Three  types  of 
quality  assessment  are  involved,  namely  fuzzy  set  theory  using  Certainty 
Factors  (CFs),  Bayesian  probability  theory  and  an  integrated  approach 
involving  also  the  positional  accuracy  of  polygon  boundaries.  In  essence,  the 
prototype  is  intended  to  provide  a  user-friendly  graphical  interface  to  assist 
ILWIS  users  to  assess  the  quality  of  their  data  and  information,  and  to 
visualize  the  results  of  the  assessment  using  a  wide  variety  of  techniques. 

Standard  ILWIS  functionality  provides  the  established  tools  for  examining 
data  and  analysing  quality,  e.g  confusion  matrices,  frequency  graphs.  PUIUS 
integrates  these.  The  graphical  user  interface  has  a  main  menu  to  allow  the 
user  to  opt  to  investigate  data  quality  or  model  quality,  and  also  to  select  an 
appropriate  visualization  technique.  As  already  indicated,  the  data  quality 
module  has  been  partially  implemented  in  the  prototype,  the  model  quality 
modile  not.  The  visualization  module  allows  a  wide  choice  of  techniques, 
including  tables,  graphs,  maps  and  oblique  views:  for  example  to  perform  a 
spatial  analysis  and  simultaneously  display  the  result.  The  associated  quality 
information  can  be  visualized  by  one  of  the  available  methods,  or  it  can  be 
accessed  in  numerical  form  by  clicking  on  a  polygon  or  pixel. 


194 


In  the  case  of  Certainty  Factors,  visualization  was  applied  to  the  quality  of  a 
crop  suitability  map.  As  the  default  representation,  four  suitability  classes, 
very  suitable  to  not  suitable,  are  represented  by  four  hues  (green,  yellow, 
orange,  red).  The  lower  the  CF,  the  greyer  (i.e.  lower  saturation)  the  hues 
become.  The  user  has  control  over  hue  choice  and  the  rate  of  decreasing 
saturation.  If  desired,  the  visual  variable  value  (lightness)  can  be  used  to 
represent  uncertainty  instead  of,  or  as  well  as,  saturation. 

In  the  case  of  quality  assessment  using  probabilities,  PUIUS  allows  the 
adjacent  display  of  results  with  or  without  using  prior  probability.  As  for  the 
CF  case  described  above,  the  visual  variables  saturation  and/or  value  can  be 
used  to  represent  uncertainty.  Also,  for  example  for  a  land  use  map  resulting 
from  the  classification  of  satellite  data,  blinking  can  be  used  to  highlight 
particular  pixels.  Several  criteria  can  be  selected,  e.g.  those  with  classification 
accuracy  above  or  below  a  specified  probability  threshold,  those  for  which  the 
classification  changes  if  prior  probability  is  applied. 

IDRISIS  INFORMATION  QUALITY  PROCESSING  TOOLS 

Apart  from  the  above-reported  work  carried  out  in  the  ILWIS  environment 
the  only  other  off-the-shelf  GIS  which  (as  the  authors  are  aware)  appears  to 
give  consideration  to  matters  of  information  quality  is  Idrisi  ( Idrisi,  1 995 ). 
Members  of  the  group  were  able  to  examine  this  in  1995.  This  is  a  developing 
system  so  findings  relate  to  a  situation  which  may  now  have  been  superseded. 

The  investigation  consisted  of  the  execution  of  two  different  spatial 
information  generation  tasks  using  Idrisi  and  analysis  of  any  uncertainty 
processing  carried  out  during  their  execution.  In  the  case  of  one  test  Idrisi 
handled  the  uncertainty  in  a  near  complete  manner  (the  location  of  a  new 
municipal  dump  -  essentially  a  buffering  operation)  and  in  the  case  of  the 
other  (identification  of  agroclimatic  zones  based  on  computed  environmental 
characteristics)  some  software  development  was  needed  (Du,1995). 

Idrisi  allows  the  insertion  of  either  a  or  probability  values  for  spatial  data. 
If  these  values  have  been  inserted  for  some  processing  modules,  error  will  be 
propagated.  The  user  has  to  know  what  values  to  insert.  We  chose,  as  far  as 
possible,  to  use  data  quality  values  which  are  standard  for  medium  scale 
topographic  data  as  supplied  in  PR  China.  There  follow  brief  descriptions  of 
Idrisi  modules  which  address  uncertainty.  (These  are  discussed  further  in 
(Idrisi,  1995)  under  the  title  "Uncertainty  and  Risk"  and  "Error  Propagation 
Formulas"  where  some,  perhaps  over-,  simplified  formulae  for  variance 
propagation  are  presented.  These  formulae  do  not  seem  to  be  implemented 
directly,  but  having  been  presented  the  users  can  implement  them  themselves.) 

PCLASS  is  an  uncertainty  handling  version  of  a  density  slicing  module. 
With  PCLASS  for  some  newly  generated  classes  (e.g.  <2000m,  <4000m, 


195 


<6000m  etc.)  a  probability  of  each  element  in  the  DTM  falling  in  that  new 
class  is  calculated,  and  stored  in  a  displayable  probability  image,  from  user 
supplied  information  on  the  quality  of  the  original  DTM  given  as  a  standard 
deviation  value.  The  central  limit  theorem  is  applied. 

BAYES  allows  the  user  to  exploit  Bayesian  Probability  theory.  The  user 
can  exploit  not  only  the  quality  of  the  datasets  being  used  in  the  analysis  (e.g. 
the  standard  deviation  of  river  or  road  coordinates  processed  through  PCLASS 
to  give  a  probability  image)  but  also  the  quality  of  the  processing  model  (e.g. 
that  the  particular  rule  for  identifying  suitable  new  landfill  dump  sites  only 
works  54%  of  the  time)  and  other  (a  priori)  information  (e.g.  an  older  study 
identifying  potential  landfill  dump  sites  in  the  same  area).  The  usefulness  of 
this  tool  is  enhanced  because  in  the  absence  of  a  priori  information  one  can 
still  exploit  the  module  (although  the  procedure  would  no  longer  be  strictly 
Bayesian)  and  at  least  use  the  information  available  on  data  quality. 

FUZZY  allows  an  image  of  uncertainty  values  to  be  generated.  These 
uncertainty  values  are  generated  by  one  of  three  different  types  of  membership 
function  (unlike  PCLASS  which  only  uses  central  limit  theorem).  Subsequent 
processing  of  the  FUZZY  images  must  be  performed  using  Idrisi's  OVERLAY 
module,  using  the  appropriate  FUZZY  Logic  rules  (i.e.  that  the  minimum 
available  certainty  factor  is  assigned  to  the  output  image  in  the  case  of 
intersection,  that  the  maximum  certainty  factor  is  assigned  in  the  case  of 
union,  etc.).  The  assignment  of  certainty  factors  (via  the  different  membership 
functions)  can  be  based  on  the  experience  or  knowledge  of  the  user  -  rather 
than  more  objective  information,  such  as  standard  deviation. 

ERRMAT  allows  the  generation  of  an  error  matrix  (sometimes  called  a 
confusion  matrix  in  Remote  Sensing  literature)  showing  the  agreement 
between  the  image  (e.g.  landuse  derived  from  the  classification  of  a  satellite 
image)  and  the  'truth'  (perhaps  obtained  by  a  set  of  ground  observations  and 
used  as  check  points). 

MCE  (Multi-Criteria  Evaluation)  supports  information  generation  using 
several  sets  of  input  data  and  user  input  processing  models.  If  standard 
deviation  values  (called  RMS  values  in  Idrisi)  are  stored  in  the  documentation 
file  of  each  input  data  set,  then  Idrisi  will  carry  out  a  simplified  variance 
propagation  to  give  an  estimate  of  output  error,  storing  it  in  the  relevant 
document  file.  The  simplification  is  such  that  the  results  imply  no  variation  in 
error  across  the  resulting  image,  which  is  neither  a  helpful  nor  a  valid 
assumption  when  the  main  purpose  of  uncertainty  analysis  is  to  guide  the 
decision  maker  to  geographic  areas  of  greater  certainty.  A  more  rigorous 
approach  has  been  implemented  by  the  authors. 

Other  Idrisi  modules  contribute  to  the  handling  or  understanding  of 
uncertainty  and  include:  REGRESS  -  for  linear  regression  analysis;  and 
RANDOM  -  for  simulating  error. 


196 


CONCLUSIONS 


The  academic  community  investigating  quality  aspects  of  GIS  is  large.  This 
group  of  authors  has  regarded  the  processing  of  uncertainty  as  a  GIS  sub- 
system. Others  may  regard  it  as  a  GIS  analytical  tool;  as  with  many  GIS 
analytical  tools  it  may  never  be  as  well  developed  in  an  off-the-shelf  system  as 
in  a  system  developed  by  specialists  particularly  interested  in  error.  However 
this  group  has  been  characterised  by  individuals  who  believe  that  the  standard 
(i.e.  off-the-shelf  tools)  should  be  as  good  as  possible  and  this  is  reflected  in 
the  fact  that  all  developments  have  taken  place  in  well-known  environments 
(AutoCAD+dBASE;  ILWIS;  Arc/Info;  Idrisi).  Obviously  development  of  error 
analysis  tools  must  continue  to  take  place  without  the  constraints  of  off-the- 
shelf-systems,  but  it  should  always  be  possible  to  implement  the  resulting 
tools  in  such  systems  -  if  they  are  to  have  any  effect  in  the  'real  world'. 

We  have  demonstrated  that  most  components  of  the  Uncertainty  Sub- 
System  can  be  implemented  in  ILWIS  and  the  sub-system  could  thus  become  a 
standard  product  and  marketed.  Although  Idrisi's  information  quality 
generation  procedures  are  (also)  incomplete,  they  are  marketed.  One  can 
commend  Clark  University.  By  making  users  at  least  sometimes  think  about 
information  quality  such  a  GIS  becomes  an  enhanced  decision  support  tool. 

ACKNOWLEDGEMENTS 

Professor  Du  Daosheng's  fellowship  at  the  Topographic  Science  Section, 
Glasgow  University,  in  1995,  was  funded  by  the  Royal  Society,  6  Carlton 
House  Terrace,  London.  The  efforts  of  our  supervisees  and  colleagues, 
without  whom  much  of  the  work  referred  to  in  this  review  would  have 
remained  undone,  namely,  Beshem  Ramlal,  Jorge  de  Oliveira,  Fan  Shien-ta, 
Elia  Barsoum  and  Stephen  McGinley  are  gratefully  acknowledged. 

REFERENCES 

Barsoum,  E.M.,  1995  "A  User  Interface  for  Error  Visualisation  in  ILWIS". 

report  to  the  Division  of  Cartography,  ITC,  1995. 
Blakemore,  M.  1983,  "Generalization  and  Error  in  Spatial  Data  Bases", 

Proceedings  of  AutoCarto  6,  1983,  pp.3 13-322 
Brown  A.  and  Elzakker,  C.P.J.M  van,  1993.  "The  use  of  colour  in  the 

cartographic  representation  of  information  quality  generated  by  a  GIS."  In: 

P.  Mesenburg  (ed.),  Procs.  Vol  2.  16th  International  Cartographic 

Conference,  Koln,  3-9  May  1993 
Chrisman,  N  R.  1983,  "The  role  of  quality  information  in  the  long-term 

functioning  of  a  GIS",  Proceedings  of  AutoCarto  6,  1 983,  pp.303-3 1 2. 
Drummond,  J.,  1987  "A  Framework  for  handling  error  in  GIS",  ITC  Jo.  1987/1 


197 


Drummond,  J.,  1990  "Models  and  Data  Quality  handled  in  a 

dBASE/AUTOCAD  GIS",  in  ITC  Journal  1989  3/4 
Du  Daosheng,  1995.  "Investigation  of  information  quality  processing  in  an 

off-the-shelf  GIS".  Report  prepared  for  the  Royal  Society  of  London  at 
Topographic  Science  Section,  Glasgow  University. 

Fan,  Shien-ta  1992  "Uncertainty  Subsystem:  Assessment  of  Model  Quality", 

MSc  Thesis,  ITC  1992 
Idrisi  for  WINDOWS,  Manual,  1995.  Clark  University  Grad.  School  of  Geog. 
ODA,  1983.  "Dominica  Land  Use  Planning  Project".  LRDC,  Tolworth  Tower, 

Surbiton,  Surrey,  KT6  7DY,  UK.  Report  #127. 
Oliviera,  J.P.  de  1992  "Generation  of  Data  Quality  Parameters  in  a  GIS",  MSc 

Thesis,  ITC  1992 

Ramlal,  B  and  Drummond,  J.E.,  1992  "A  Prototype  Uncertainty  Subsystem 
implemented  in  ITC's  ILWIS  PC  based  GIS  and  tested  on  a  Dutch  Land 
Reallotment  Project",  EGIS  Conference  Proceedings,  Munich  1992 

Ramlal,  Bheshem  1991  "Communicating  Information  Quality  in  a  GIS 
Environment",  MSc  Thesis,  ITC  1991 

Van  Elzakker,  C,  Ramlal,  B  and  Drummond,  J.E.,  1992  "The  Visualisation  of 
GIS  Generated  Information  Quality",  August  1992,  Archives  ISPRS 
Congress  XVI,  Commission  IV 

BIOGRAPHICAL  SKETCH 

Jane  Drummond  is  a  staffmember  of  Glasgow  University's  Topographic 
Science  Section.  Prior  to  that  she  worked  in  ITC's  Department  of 
Geoinformatics,  the  UK  NERC's  Experimental  Cartography  Unit,  UNB's 
Department  of  Geodesy  and  Geomatics  Engineering,  and  Hunting  Surveys. 

Allan  Brown  is  a  staffmember  of  ITC's  Department  of  Geoinformatics. 
Previously  he  has  worked  in  Gadja  Mada  University's  Department  of 
Geography,  Glasgow  University's  Topographic  Science  Section  as  an  Assistant 
Lecturer,  and  at  Fairey  Surveys. 

Du  Daosheng  is  a  staffmember  of  the  National  Key  Laboratory  at  Wuhan 
Technical  University  of  Surveying  and  Mapping,  holding  the  rank  of 
professor.  Prior  to  that  he  was  Head  of  the  Department  of  Cartography  at  the 
same  institute.  From  1982  -  1984  he  worked  in  the  UK  NERC's  Experimental 
Cartography  Unit. 

Corne  van  Elzakker  is  a  staffmember  of  ITC's  Department  of 
Geoinformatics.  Prior  to  that  he  was  a  lecturer  in  the  Cartography  Section  of 
the  Department  of  Geography  of  Utrecht  State  University. 


198 


Communication  of  Uncertainty  in 
Spatial  Data  to  Policy  Makers 

Morwenna  Spear1,  Jane  Hall2  and  Richard  Wadsworth2 


Abstract.  -  There  is  increasing  interest  in  the  spatial  aspects  of  many 
policy  decisions.  To  be  valid  decision-making  tools  it  is  important  that 
maps  not  only  convey  what  is  known,  but  also  where  the  uncertainties  lie. 
This  paper  reports  the  results  of  tests  conducted  to  assess  the  ability  of 
various  methods  of  error  representation  to  accurately  assist  interpretation 
of  maps.  The  work  uses  the  predicted  surface  and  associated  error  surface 
(both  generated  by  a  kriging  function)  to  test  the  ability  of  prospective  and 
current  users  to  interpret  the  data  presented,  and  to  draw  simple  logical 
conclusions  about  it.  The  ultimate  aim  is  to  study  the  effect  of  user 
perception  of  accuracy  on  the  real-life  decisions  made  using  interpolated 
spatial  data. 


INTRODUCTION 

Maps  are  an  effective  method  for  communicating  the  spatial  aspect  of 
scientific  data  to  policy  makers.  The  current  high  priority  on  environmental 
issues  means  that  maps  may  be  required  to  show  levels  of  pollution  and  their 
effects  on  different  ecosystems  to  policy  makers.  Creation  of  these  maps  usually 
requires  the  use  of  models  and  interpolation  techniques.  This  has  been  made 
easier  as  Geographical  Information  Systems  (GIS)  have  evolved  to  handle  the 
large  quantities  of  data  and  perform  elaborate  calculations  quickly.  Unfortunately 
these  developments  have  an  associated  cost,  in  that  the  new  data  processing 
methods  may  introduce  and  magnify  uncertainties  and  errors  in  ways  which 
previously  have  not  required  such  consideration.  Conversely,  the  enhanced  data 
handling  ability  equips  us  to  study  these  errors  and  uncertainties  more  easily  and 
we  therefore  should  begin  to  communicate  this  information  to  map  users. 

The  analysis  of  sources  of  errors  and  their  propagation  has  been  an  area  of 
increasing  interest  since  the  mid  1980s  (Heuvelink  et  al.  1989).  Many 
researchers  have  investigated  the  topic  and  used  various  techniques  (for 
example,  see  Veregin  1994).  However,  there  have  been  fewer  reported 
investigations  into  methods  for  displaying  values  of  error  and  uncertainty.  Maps 
may  sometimes  be  presented  with  an  estimated  error,  for  example,  digital 
elevation  models  published  by  the  United  States  Geological  Service  are  supplied 

'  Research  Assistant,  TRADA.  High  Wycombe.  Bucks.  United  Kingdom 

2  Higher  Scientific  Officer,  Institute  of  Terrestrial  Ecology,  Huntingdon.  Cambs.  United  Kingom 


199 


with  an  indication  of  their  root-mean-squared-error.  Any  report  of  error  is 
welcome,  but  as  Lee  et  al.  (1992)  point  out  the  magnitude  of  the  error  will  vary 
over  the  area  and  a  different  approach  may  be  more  appropriate.  In  presenting 
interpolated  data,  associated  error  estimates  are  sometimes  presented  in  map 
form  (Webster  et  al  1991).  This  allows  the  map  user  to  be  aware  which  regions 
within  the  area  are  likely  to  be  prone  to  high  errors,  and  to  make  a  more 
informed  decision  in  the  light  of  this  extra  knowledge. 

Work  at  the  Environmental  Information  Centre  of  the  Institute  of  Terrestrial 
Ecology  includes  presenting  maps  of  environmental  data  to  policy  makers  in 
Government  departments.  Typically  these  maps  show  levels  of  pollutants 
deposited  generated  by  interpolating  point  data  from  across  the  UK.  They  may 
then  be  combined  with  suitable  environmental  receptor  maps.  The  resulting 
maps  may  then  be  used  to  identify  areas  where  the  threshold  of  a  pollutant  is  in 
excess  of  the  ecosystem's  ability  to  buffer  it.  This  research  has  considered 
methods  for  presenting  the  uncertainty  in  interpolated  data  to  decision  makers. 
Several  methods  of  communicating  uncertainty  in  raster  data  sets  have  been 
developed  and  tested  for  their  ability  to  convey  understanding  of  error. 
Ultimately  we  hope  to  assess  the  sources  and  compounding  factors  for  all  errors 
associated  with  our  data. 

INVESTIGATION  METHODS 

The  emerging  study  area  of  error  visualisation  draws  on  knowledge  from  a 
wide  range  of  disciplines:  geostatistics,  computing,  cartography,  psychology.  In 
the  development  of  new  presentation  techniques  our  aim  was  to  inter-relate  the 
knowledge  and  understanding  of  the  various  disciplines  to  ensure  the  solution 
found  was  reasonable. 

Investigation  of  Interpolated  Data  Errors 

To  gain  a  fuller  understanding  of  the  nature  of  the  errors  present  in 
interpolated  data  we  generated  several  known  data  sets,  resembling  interpolated 
deposition  maps.  These  allowed  better  understanding  of  the  errors  which 
interpolation  introduced.  From  these  hypothetical  maps  we  sampled  z-values 
using  various  sampling  schemes.  Kriging  was  chosen  for  interpolation  of  these 
sampled  data  as  it  gave  an  accessible  measure  of  uncertainty.  Using  this 
technique  a  number  of  predicted  maps  and  predicted  error  values  were  generated. 
We  were  able  to  compare  the  estimates  with  the  real  values,  which  allowed  us  to 
verify  that  the  errors  predicted  with  the  interpolation  were  appropriate.  We  found 
that  the  predicted  error  values  were  approximately  equivalent  to  one  standard 
error;  in  that  sixty-eight  percent  of  actual  errors  were  smaller  than  the  predicted 
error.  This  implied  that  the  error  values  could  be  used  to  define  a  confidence 


200 


interval.  This  assumption  was  confirmed  by  finding  that  an  interval  twice  the 
predicted  error  included  95%  of  actual  errors.  Using  the  predicted  error  values, 
68%  confidence  intervals  were  calculated  for  each  predicted  map.  It  should  be 
noted  that  further  study  will  address  the  correct  use  of  confidence  intervals  for 
modelled  error.  This  paper  is  concerned  with  the  validity  of  the  visualisation 
methods. 

Presentation  of  Estimated  Uncertainty 

Past  tests  have  shown  that  layer  tint  maps  are  the  most  successful  method  for 
visualisation  of  a  landscape  (Phillips  et  al  1975,  Phillips  1984).  The  colour  of 
the  layers  reminds  the  map  user  of  the  value  between  the  contour  lines  when 
scanning  a  large  area  of  the  map.  The  limitations  when  visualising  a  surface 
imply  that  the  inclusion  of  uncertainty  as  an  extra  variable  will  pose  greater 
problems  for  our  visual  memory.  Layer  tint  maps  were  used  as  they  are  the  least 
demanding  method  of  display,  as  well  as  being  the  simplest  way  of  handling 
raster  data. 

The  simplest  way  of  presenting  the  estimated  uncertainty  is  as  a  separate  map. 
However  the  definition  of  a  confidence  interval  enables  us  to  communicate  the 
errors  present  in  an  entirely  different  manner.  The  use  of  raster  data  sets  allows 
the  uncertainty  and  confidence  interval  to  be  calculated  on  a  cell-by-cell  basis. 
The  confidence  interval  can  be  presented  in  three  separate  maps  (being  the 
predicted  map  and  the  upper  and  lower  limits  of  the  interval  -  see  figure  2),  or 
the  three  maps  can  be  combined  into  one  image  (figure  3).  Either  form  of 
presentation  should  be  accompanied  by  a  statement  of  the  level  of  confidence 
presented.  Using  the  three  maps  separately,  map  users  can  be  confident  in  their 
extraction  of  information.  For  example,  the  map  user  may  check  that  they  have 
selected  the  correct  end  of  the  confidence  interval  for  their  current  use.  Where 
the  maps  have  been  combined,  the  process  gives  rise  to  hatched  areas  where  the 
classes  from  each  of  the  three  constituent  maps  overlap.  The  hatchings  give  an 
impression  of  the  uncertainty  of  the  information  being  presented,  marking  the 
extent  to  which  the  class  boundary  may  differ  in  reality.  It  is  possible  to  visualise 
the  effect  of  the  uncertainty  in  all  three  dimensions,  without  the  need  to  cross 
refer  between  maps. 

SURVEY  METHODOLOGY 

Having  developed  several  visualisation  methods  the  next  phase  of  the 
investigation  was  to  conduct  a  survey.  This  was  designed  to  assess  the  level  of 
information  which  could  be  conveyed  about  uncertainty  in  interpolated  maps  and 
test  a  representative  selection  of  the  visualisation  techniques  developed.  The 
composition  of  tests  which  have  been  conducted  in  the  past  (Phillips  et  al.  1975, 


201 


Phillips  1982,  Potash  et  at.  1978)  were  considered,  and  have  been  modified  to 
reflect  the  requirements  of  presenting  error  to  decision  makers.  We  wished  to 
check  that  information  from  the  predicted  map  was  still  accessible,  but  more 
importantly  to  assess  the  ability  of  candidates  to  determine  the  effects  of 
uncertainty  on  that  prediction. 

The  three  techniques  selected  to  be  investigated  in  our  survey  were 
presentation  of  the  error  on  a  map  beside  the  predicted  map  (figure  1); 
presentation  of  a  confidence  interval  as  three  separate  maps  (figure  2)  and 
presentation  of  a  confidence  interval  as  one  combined  map  (figure  3) 


Figure  1.  -  Presentation  of  interpolated  prediction  (left)  and  the  predicted  error  (right). 


Figure  2.  -  Presentation  of  interpolated  prediction  (centre)  with  the  confidence  interval 
limits;  lower  (left)  and  upper  (right). 

Structure  of  Survey 

The  survey  consisted  of  two  sections:  assessment  of  users'  skill  and 
assessment  of  users'  preference. 

The  first  section  assessed  the  candidate's  ability  to  understand  maps  presented 


202 


0        10      20      30      40  50 

Figure  3.  -  Presentation  of  the  confidence  interval  maps  as  a  combined  image. 

using  the  three  techniques  listed  above.  Some  maps  may  be  used  to  locate  areas 
in  which,  for  example,  pollutant  deposition  values  exceed  an  acceptable  limit. 
One  set  of  questions  addressed  this  issue,  asking  the  candidates  to  mark  the  areas 
in  which  the  values  might  exceed  a  given  level  (when  the  error  was  taken  into 
account).  A  second  type  of  question  asked  the  candidates  to  mark  the  highest  and 
lowest  points  on  the  map,  to  assess  whether  they  could  understand  the  key  and 
layer  tint  used.  Both  question  types  were  repeated  for  all  three  methods,  as  well 
as  for  a  control  question  in  which  no  error  was  presented.  A  further  type  of 
question  asked  the  candidates  to  identify  areas  where  additional  data  samples 
were  needed  to  reduce  the  resultant  error. 

Early  in  the  first  section,  where  the  methods  had  been  presented  but  not 
explained,  the  candidates  were  asked  for  their  opinion  of  the  methods.  This  was 
to  gain  an  impression  of  their  immediate  reaction,  simulating  a  situation  where  a 
new  method  of  presentation  was  encountered  in  a  journal  or  book  without 
detailed  explanation.  The  opportunity  to  give  a  more  detailed  response  on  the 
three  methods  was  given  at  the  end  of  the  test  after  the  candidate  had  had  further 
experience  with  all  three  methods. 

The  second  section  contained  questions  which  allowed  the  candidates  to 
select  the  method  with  which  they  "felt  most  comfortable"  to  answer  questions 
similar  to  those  described  above.  Here,  all  three  methods  were  given,  on  the 
same  page,  using  maps  of  equal  size.  The  candidates  were  not  limited  to  the  use 
of  only  one  method  but  were  asked  to  tick  each  map  which  they  used.  These 


203 


questions  were  used  to  assess  the  method  the  candidates  preferred,  and  hence  the 
method  most  likely  to  be  popular  in  a  real  situation.  The  information  was  also 
used  to  check  that  the  most  popular  method  was  actually  the  method  which  was 
used  correctly. 

Candidates  were  selected  to  represent  as  wide  a  selection  of  potential  map 
users  as  possible.  In  all,  24  science  and  administrative  staff  participated, 
representing  a  range  of  mapping  expertise  and  seniority.  The  level  of 
geographical  qualification  held  by  each  candidate  varied  from  none  to 
postgraduate.  The  frequency  of  map  use  varied  from  never  to  daily.  The 
candidates  were  told  that  completing  the  questionnaire  would  take  approximately 
30  minutes,  but  that  if  they  felt  they  required  longer,  or  needed  to  finish  sooner 
in  order  to  meet  their  other  work  commitments,  they  could  do  so. 

ANALYSIS  OF  RESULTS 

The  candidates'  responses  have  been  used  to  investigate  many  aspects  of  the 
presentation  techniques.  Most  important  is  the  comparison  between  the  methods 
using  error  maps  with  methods  showing  confidence  intervals.  The  comparison 
between  the  two  confidence  interval  presentations  chosen  is  also  interesting,  but 
we  anticipate  future  developments  in  the  presentation  of  confidence  intervals 
will  lead  to  altered  preferences.  The  results  outlined  below  should  be  a  starting 
point  for  further  investigations. 

Time  and  Experience 

Before  detailed  analysis  of  the  results,  it  was  necessary  to  consider  the  effect 
of  external  factors  on  the  candidates'  responses.  As  was  to  be  expected,  the 
candidates'  levels  of  geographical  qualification  and  their  frequency  of  map  use 
were  found  to  be  related  to  their  performance  in  the  survey.  However,  the  wide 
range  of  geographical  skills  which  civil  servants  and  policy  makers  possess, 
made  it  desirable  to  ensure  that  this  range  of  skill  was  also  present  in  the  survey. 
The  requirement  for  sufficient  explanation  of  the  occurrence  and  effects  of 
uncertainty  in  maps  cannot  be  understated  as  we  must  be  sure  that  the  topic  is 
understood  correctly. 

It  was  also  important  to  assess  the  effect  of  the  time  taken  to  complete  the 
questionnaire.  The  quality  of  answer  was  shown  to  have  no  correlation  with  the 
amount  of  time  spent  in  responding.  So  the  lack  of  imposed  time  constraint 
appears  to  have  had  no  biasing  effect  on  the  results  of  the  survey. 

Identification  of  Threshold  Values 

The  responses  were  scored  by  two  methods.  Initially  they  were  assessed  with 


204 


reference  to  the  ideal  response.  A  score  was  allocated  to  reflect  the  level  to 
which  the  candidate  had  understood  and  implemented  the  idea.  This  indicated 
that  the  combined  map  gave  the  highest  score  for  most  candidates.  This  scoring 
method  was  quite  arbitrary,  so  a  second  method  was  designed  to  be  more 
quantitative.  A  regular  grid  of  points  was  prepared  on  transparency,  each  point 
being  marked  as  being  above  or  below  the  threshold.  The  grid  was  overlaid  on 
the  response,  and  the  identified  points  split  into  four  categories:  correctly 
marked,  correctly  unmarked,  incorrectly  marked  and  incorrectly  unmarked.  The 
counts  were  used  to  conduct  chi-squared  tests  which  identified  the  responses 
which  were  wrong.  The  percentage  of  candidates  giving  correct  responses  was 
used  as  an  indicator  of  the  success  of  the  method.  Occasionally  a  candidate  had 
not  answered  the  question  or  had  marked  the  map  in  a  way  which  was  not  suited 
for  this  type  of  analysis,  so  these  responses  have  been  treated  as  a  separate 
category.  The  results  for  each  presentation  method  are  shown  in  figure  4. 


no  uncertainly        predicted  pair         predicted  pair         CI  three  imps         CI  three  mips  □combined  Ocorrfimcd 

method  of  visualisation 


Figure  4.  -  Proportion  of  candidates  identifying  exceeded  areas  correctly.  This  question 
type  was  used  once  where  no  uncertainty  was  presented,  then  repeated  twice  for  each  of 
the  three  uncertainty  display  methods. 

The  most  visible  trend  was  that  the  use  of  the  predicted  error  map  alongside 
the  predicted  map  gave  a  significantly  lower  standard  of  result  than  where  either 
form  of  confidence  interval  was  used.  Indeed,  with  the  confidence  interval 
presentations  the  scores  were  similar  to  the  control  example  where  no  error  was 
presented  with  the  map.  In  situations  where  the  map  user  requires  a  visual 
perception  of  the  extremes  which  the  uncertainty  implies,  the  confidence  interval 
is  the  more  useful  tool.  The  results  also  indicate  that  the  combined  map  of  the 
confidence  interval  was  the  most  accurately  used.  This  observation  might  be  an 
artefact  of  the  relatively  small  number  of  questions  asked.  Ideally  further  studies 
would  be  conducted  to  assess  the  most  effective  method  amongst  the  possible 
confidence  interval  representations.  Other  methods  for  presenting  the  confidence 
interval  are  currently  being  developed. 


205 


Assessing  Location  For  Extra  Samples 


Where  the  candidates  were  requested  to  locate  areas  where  extra  samples 
were  needed,  the  majority  of  candidates  opted  to  use  the  predicted  error  map. 
There  were  however,  two  other  methods  which  candidates  justifiably  chose  to 
use.  One  was  to  refer  to  both  the  predicted  error  and  the  original  map.  This 
reduced  the  uncertainty  in  areas  where  it  was  high  and  verified  the 
"topographical"  features  which  the  map  seemed  to  predict.  The  other  was  to  use 
the  combined  map  to  locate  areas  where  the  cells'  confidence  interval  spanned 
two  or  more  class  intervals.  This  was  a  very  interesting  option,  as  the  tint  layers 
would  ideally  be  selected  to  reflect  the  classes  in  which  the  map  user  was 
interested.  Areas  which  were  borderline  between  two  classes  would  be  the  areas 
in  which  the  uncertainty  would  have  the  most  influence  over  that  decision. 
Clarifying  the  boundaries  in  these  areas  would  be  the  most  cost  effective  action. 

Candidates'  Preferences 

Where  candidates  were  allowed  to  choose  between  the  three  methods  of 
presentation,  the  most  popular  method  was  the  three  separate  confidence  interval 
maps,  but  the  use  of  the  combined  map  came  a  close  second.  Some  candidates 
opted  to  use  both  of  these  methods,  and  explained  this  choice,  saying  that  the 
three  separate  maps  were  useful  for  ensuring  that  they  selected  the  correct 
information  and  map.  The  combined  map  was  additionally  useful  in  spatially 
relating  the  three  maps  to  the  one  area  under  consideration  and  identifying  the 
areas  on  the  original  map. 

The  candidates'  opinions  expressed  in  the  free  response  questions  often 
indicated  quite  clearly  their  preferences  between  methods.  However  a  number  of 
candidates  made  no  comment.  The  trend  for  preferring  the  confidence  interval 
was  explained  by  the  difficulty  of  superimposing  the  original  two  maps,  and  the 
simplicity  of  using  the  confidence  interval.  The  reasons  expressed  for  lack  of 
conviction  in  the  combined  confidence  interval  were  problems  with  the 
explanation,  difficulties  with  the  key  and  greyscale,  and  the  presence  of  too 
much  information.  However  a  number  of  people  expressed  a  preference  for  this 
method,  as  the  shading  makes  it  easier  to  visualise  patterns. 

CONCLUSIONS 

Currently  the  presentation  of  uncertainties  and  errors  in  maps  is  very  limited. 
This  situation  has  perpetuated  many  untrained  users'  perceptions  that  maps  are 
infallible.  Where  maps  are  subject  to  errors  and  uncertainties  which  might 
influence  policy  decisions,  these  now  can  and  should  be  presented.  The  results  of 


206 


the  survey  presented  here  show  that  it  is  possible  to  convey  information  on 
uncertainty  and  error,  even  to  people  who  are  unfamiliar  with  the  data.  The  use 
of  a  confidence  interval  gives  an  enhanced  understanding  of  the  potential  effect 
of  the  uncertainty  on  the  reported  values.  There  are  many  potential  methods  for 
presenting  a  confidence  interval,  the  two  which  have  been  reported  here  have 
been  investigated  in  depth,  but  others  have  also  been  considered  by  the  authors. 
Other  methods  will  doubtless  be  developed  to  satisfy  specific  requirements. 

The  presentation  of  the  error  as  a  map  in  its  own  right  is  useful  in  the 
planning  stages  of  a  sampling  scheme  for  interpolation,  or  where  a  model  needs 
to  be  improved.  However,  the  use  of  the  combined  confidence  interval  map  has 
proved  more  effective  in  targeting  areas  where  the  uncertainty  has  the  greatest 
effect  on  the  data  currently  under  consideration.  The  use  of  a  confidence  interval 
can  be  employed  in  any  raster  data  set,  and  the  confidence  interval  can  be 
defined  according  to  the  requirements  of  the  users  and  the  quality  of  the  data. 
The  use  of  a  confidence  interval  in  mapping  is  a  relatively  new  concept  which 
has  the  possibility  of  being  a  major  advancement  in  understanding  uncertainty 
within  maps.  As  information  on  uncertainty  in  critical  loads  and  deposition 
estimates  becomes  available  the  communication  of  uncertainty  to  policy  makers 
will  become  a  standard  procedure. 

REFERENCES 

Heuvelink  G.B.M.,  Burrough  P.A.  and  Stein  A.  1989  Propagation  of  errors  in 

spatial   modelling   with   GIS   International      Journal   of  Geographical 

Information  Systems  3(4)  pp. 303-322 
Lee  J.,  Snyder  P.K.  and  Fisher  P.F.  1992  Modelling  the  effect  of  data  errors  on 

feature    extraction    from    Digital    Elevation    Models    Photo grammetric 

Engineering  and  Remote  Sensing  58(10)  pp.  1461- 1467 
Phillips  R.J.,  DeLucia  A.  and  Skelton  N.  1975  Some  objective  tests  of  the 

legibility  of  relief  maps  Cartographic  Journal  12(1)  pp. 39-46 
Phillips  R.J.  1982  An  experimental  investigation  of  layer  tints  for  relief  maps  in 

school  atlases  Ergonomics  25(12)  pp.1 143-1 154 
Potash  L.M.,  Farrell  J. P.  and  Jeffrey  T.  1978  A  technique  for  assessing  map 

relief  legibility  Cartographic  Journal  15(1)  pp. 28-35 

Veregin  H.  1994  Integration  of  simulation  modelling  and  error  propagation  for 
the  buffer  operation  in  GIS  Photogrammetric  Engineering  and  Remote 
Sensing  60(4)  pp.427-435 

Webster  R.,  Campbell  G.W.  and  Irwin  J.G.  1991  Spatial  analysis  and  mapping 
the  annual  mean  concentrations  of  acidity  and  major  ions  in  precipitation  over 
the  United  Kingdom  in  1986  Environmental  Monitoring  and  Assessment 
16(1)  pp.1-17 


207 


Measuring  the  Performance  of 
Algorithms  for  Generating 
Ground  Slope 

William  H.  Ryder1  and  Demetra  E.  Voyadgis2 


Abstract.  -  Calculation  of  ground  slope  is  a  fundamental  geographic 
information  system  (GIS)  utility.  While  accurate  slope  prediction  is 
acknowledged  to  be  a  fundamental  capability,  reality  is  that  users  normally 
apply  the  slope  utility  in  their  GIS  without  considering  accuracy  of  the 
product.  As  the  Army's  center  of  expertise  for  digital  topographic  data,  the 
Digital  Concepts  and  Analysis  Center  at  the  U.S.  Army  Topographic 
Engineering  Center  (TEC)  has  conducted  a  study  of  the  accuracy  of  several 
slope  generation  algorithms.  This  study  was  designed  to  field  test  existing 
slope  generation  algorithms,  to  identify  one  or  more  that  performed 
satisfactorily,  and  to  recommend  incorporation  of  those  into  Army  systems. 
Several  Army  source  algorithms  as  well  as  algorithms  embedded  in 
commercial  off  the  shelf  (COTS)  software  packages  were  tested.  Rapid 
field  collection  of  point  slope  data,  a  key  aspect  of  this  study,  was  made 
possible  through  Global  Positioning  System  (GPS)  and  laser  range  finding 
technology.  To  assess  slope  algorithm  accuracy,  field  slope  measurements 
were  compared  with  computer  generated  slope  values. 


INTRODUCTION 

Detailed  mapping  of  the  Earth's  surface  configuration  (slope)  has  always  been 
time  intensive.  Traditional  slope  compilation  methods  are  very  labor  intensive, 
requiring  analysis  of  contour  line  information  by  a  human  interpreter.  Compilation 
time  can  be  dramatically  reduced  by  using  a  computer  to  generate  slope  from  digital 
elevation  data.  Key  ingredients  in  this  process  are  elevation  data  with  adequate 
resolution  and  an  algorithm  that  performs  satisfactorily.  This  study  was 
commissioned  to  examine  algorithms  for  slope  generation  from  elevation  data. 

Automated  slope  generation  has  been  a  topic  of  research  for  many  years.  Efforts 
by  the  Defense  Mapping  Agency  (DMA)  and  the  U.S.  Army  Topographic 
Engineering  Center  (TEC)  during  the  1970's  were  abandoned  due  to  the  lack  of 
accurate  Digital  Terrain  Elevation  Data  (DTED).    Since  then,  source  data  has 


Physical  Scientist.  U.S.  Army  Topographic  Engineering  Center.  Alexandria.  VA. 


-  Geographer.  L.S.  Army  Topographic  Engineering  Center.  Alexandria.  VA. 

208 


improved  and  new  algorithms  have  been  written.  Problems  also  existed  because 
users  wanted  a  product  that  resembled  a  traditional,  manually  compiled  slope  overlay 
and  slope  generated  from  gridded  elevation  data  was  blocky  and  disjointed.  This  is 
less  of  a  problem  today  because  slope  is  mainly  used  in  computer  models  and  seldom 
as  a  stand-alone  product.  With  higher  spatial  resolution  elevation  data  and  faster 
computers,  automated  slope  generation  is  a  reality.  The  question  remains,  given 
identical  elevation  data  inputs,  which  algorithm  creates  the  best  representation  of  the 
Earth's  surface  configuration? 

APPROACH 

The  approach  used  by  this  study  was  to  review  literature  on  algorithms  used 
currently,  and  to  select  candidates  based  on  the  availability  and  acceptance  by  the 
users.  Three  grid-based  algorithms  were  examined.  Also  examined  were  two 
manually  compiled  slope  overlays.  (See  Table  1). 

Table  1.— Slope  Generation  Methods. 


System/Author  Method  used 


ARC  INFO 

Uses  the  average  maximum  technique. 

Calculates  slope  from  the  third  order  finite 

difference  method  as  described  by  Horn, 

1969.  Doubles  the  weight  of  the  nearest 

neighbors  in  the  cardinal  directions. 

ERDAS  IMAGINE 

Similar  to  ARC/INFO,  but  weighs  all 

neighbors  equally. 

U.S.  Army  Waterways  Experiment  Station 

Calculates  slope  from  normal  vectors  and  the 

(WES) 

second  order  finite  difference  method. 

1:24.000  Slope  Overlay 

Manually  compiled  from  the  contours  of  a 

1:24,000  USGS  Quadrangle. 

Tactical  Terrain  Analysis  Data  Base  Slope 

Manually  compiled  by  DMA  from  the 

Overlay 

contours  of  a  1:50,000  Topographic  Line 

Map. 

The  algorithms  were  tested  using  a  high  spatial  resolution  elevation  data  set. 
Elevation  postings  for  this  data  set  were  collected  at  five  meter  spacings.  Besides 
generating  slope  using  the  five  meter  data,  spacings  were  thinned  to  30  and  100 
meters.  This  was  done  to  approximate  the  elevation  post  spacing  of  DTED  Levels 
2  and  1  respectively. 

The  slope  values  that  resulted  from  the  different  algorithms  and  resolutions  were 
categorized  according  to  the  ranges  used  on  a  DMA  Tactical  Terrain  Analysis  Data 
Base  (TTADB)  (See  Table  2).  An  TTADB  is  a  collection  of  six  overlays  (obstacles, 


209 


soils,  slope,  surface  drainage,  transportation  and  vegetation)  used  by  the  military  for 
tactical  planning.  The  categorized  results  were  compared  statistically  against  field 
data  to  calculate  the  overall  accuracy. 


Table  2.--TTADB  Slope  Categories 


Slope  category  Range  (percent  slope) 

A  0<  3 

B  >3  <  10 

C  >10  <  20 

D  >20  <  30 

E  >30  <  45 

F  >45 


DATA 
Data  Used 

To  meet  study  objectives,  data  used  had  to  be  accurate  and  have  high  spatial 
resolution.  Spatial  resolution  had  to  be  at  least  equal  to  that  of  DTED  Level  2  (  30 
meters).  Data  that  met  these  requirements  were  available  over  the  Yakima  Training 
Center,  Washington.  An  elevation  matrix  was  created  on  a  digital  stereo 
photogrammetric  workstation  using  GPS  controlled  9"  by  9"  imagery. 

The  Yakima  Training  Center  was  chosen  for  this  study  because  of  the  variety  of 
slope  categories  found  there.  Additionally,  the  road  network  throughout  the  area 
allowed  for  easier  access  to  site  locations.  To  conserve  computer  resources,  a  four 
kilometer  by  four  kilometer  area  was  used  as  the  test  area.  This  area  contained  the 
required  variety  of  terrain/slope  categories. 

Collection  of  Field  Data 

The  procedure  for  collection  of  data  consisted  of  two  parts,  test  point  selection  and 
slope  measurement  in  the  field. 

Test  Point  Selection 

Field  objectives  of  this  study  required  that  each  point  visited  in  the  field  meet  the 
following  conditions:  points  had  to  be  separated  by  at  least  125  meters  (this  was  to 
ensure  that  the  longest  distance  measured,  100  meters,  did  not  fall  into  area  covered 
by  other  points),  and  the  125-meter  radius  circle  surrounding  the  point  had  to  be 
totally  contained  within  a  slope  category,  not  crossing  over  into  another  category. 
To  meet  these  conditions,  approximately  50  points  from  each  slope  category  were 
preselected  on  the  1 :24.000  slope  overlay.  The  actual  number  of  points  visited  was 
lower  because  of  firing  range  restrictions  at  Yakima. 


210 


Once  in  the  field,  personnel  navigated  to  the  preselected  coordinates  using  a 
Precision  Lightweight  Global  Positioning  System  Receiver  (PLGR).  When  using  the 
military  key,  the  PLGR  is  accurate  to  within  ten  meters.  This  amount  of  accuracy  is 
sufficient  for  measuring  slope  values  at  thirty  and  one  hundred  meter  post  spacings, 
but  higher  accuracies  were  needed  to  measure  slope  at  five  meter  spacings.  To 
achieve  higher  accuracies,  differential  GPS  was  used.  After  reaching  a  point  using 
the  PLGR,  readings  were  taken  using  the  differential  GPS  unit.  This  unit  took 
position  readings  every  five  seconds  until  sixty  positions  were  recorded.  These 
positions  were  later  differentially  corrected  by  comparing  them  with  readings  taken 
by  a  GPS  base  station  set  up  over  a  known  point.  The  average  of  the  corrected 
readings  provided  a  coordinate  that  was  accurate  to  approximately  two  and  one-half 
meters. 

Field  Measurement  of  Slope  Values 

Slope  was  measured  in  the  field  using  laser  ranging  equipment.  Once  a  point  was 
found  using  the  PLGRs,  and  its  actual  position  was  recorded  with  the  differential 
GPS.  one  analyst  remained  on  the  point  while  the  other  analyst  paced-off  a  distance 
of  five  meters  in  the  direction  of  the  steepest  slope.  The  exact  distance  was  measured 
using  the  range  gun,  and  the  analyst's  position  was  adjusted  accordingly.  When 
using  the  range  guns,  care  was  taken  to  ensure  that  the  measurement  point  on  the 
analyst  was  at  the  same  height  as  the  gun  itself.  Maximum  slope,  when  not  readily 
apparent,  was  determined  by  measuring  in  several  directions.  The  laser  ranging 
equipment  is  ideal  for  this  work,  instantly  giving  slope  percent  and  horizontal 
distance  as  well  as  azimuth.  After  recording  the  readings,  the  analyst  moved  to  thirty 
meters  from  the  start  point,  again  making  sure  to  be  on  the  maximum  slope,  and  the 
measurements  were  taken.  This  procedure  was  repeated  at  1 00  meters  from  the  start 
point.  Maximum  slope  was  determined  for  each  distance.  The  maximum  was  not 
always  in  the  same  direction  for  all  distances. 

MANUALLY  GENERATED  SLOPE  METHODOLOGY 

The  1 :24.000  slope  overlay  was  manually  compiled  in-house,  using  the  contours 
from  a  United  States  Geological  Survey  1:24,000  topographic  map  and  a  slope 
wedge.  The  slope  wedge  uses  a  series  of  circles  whose  sizes  correspond  to  the  slope 
categories.  The  category  is  determined  by  the  largest  circle  that  can  fit  between  two 
contours.  As  the  distance  between  the  two  contours  increases  or  decreases,  the  slope 
category  changes.  A  minimum  polygon  size  of  250  meters  by  250  meters  (ground 
distance)  was  used  in  accordance  with  DMA  specifications.  This  is  the  same  method 
that  DMA  used  to  compile  the  TTADB  from  the  contours  of  a  1 : 50,000  Topographic 
Line  Map. 


211 


COMPUTER  GENERATED  SLOPE  METHODOLOGY 


Three  methods  were  used  for  generating  the  slope  thematic  maps:  an 
ARC/INFO  method,  a  WES  method,  and  an  ERDAS  IMAGINE  method. 

The  ARC/INFO  Method 

A  matrix  of  the  elevations  points  was  output  in  ASCII  format.  The  ARC 
generate  command  was  used  to  generate  a  point  coverage,  which  was  brought  into 
the  GRID  module  through  the  pointgrid  function.  Next,  the  field  locations  were 
downloaded  from  the  differential  GPS  in  ASCII  format.  Again,  the  generate 
command  was  used  to  create  a  point  coverage,  the  pointgrid  command  used  to 
create  a  grid  coverage. 

The  original  elevation  data  set,  at  5 -meter  spacing,  was  thinned  to  30  meters  and 
1 00  meters  using  the  resample  command.  For  each  of  these  three  data  sets,  the 
following  procedures  were  performed: 

•  define  the  projection 

•  calculate  the  slope 

•  classify  the  data 

•  extract  the  slope  category  values  for  the  field  locations. 

The  projectdefine  command  was  used  to  set  the  projection,  datum,  and  units  of 
measure  to  match  those  used  during  data  extraction.  (This  was  also  performed  on 
the  field  location  data  set.)  The  slope  was  calculated  using  the  slope  command.  A 
remap  table  of  values,  using  the  TTADB  slope  categories,  was  created  and  used 
with  the  reclass  function  to  produce  the  thematic  map.  Finally  the  combine 
command  was  used  to  extract  the  slope  category  values.  The  combine  command 
combines  two  or  more  coverages  and  outputs  a  coverage  of  unique  values.  The 
database  table  associated  with  the  resulting  coverage  contains  the  values  from 
each  of  the  original  coverages.  Therefore,  when  the  reclassified  slope  grid  and  the 
field  location  grid  are  input  to  the  combine  function,  the  resulting  grid  coverage 
contains  only  those  grid  cells  that  are  common  to  both  input  grids.  The  database 
table  for  the  resulting  grid  contains  the  slope  category  value  and  the  field  location 
number.  These  numbers  were  then  used  in  the  statistical  comparisons. 

The  WES  Method 

The  WTS  algorithm  is  a  stand-alone  utility  that  generates  slope  values  directly 
from  a  matrix  of  elevation  points,  and  returns  the  results  as  a  matrix  in  ASCII 
format.  Therefore,  a  software  suite  was  needed  to  classify  the  slope  values  and  to 
extract  the  values  for  the  field  locations.  ARC/INFO  was  used  for  these  purposes. 
Using  the  ARC/INFO  commands  stated  above,  the  following  procedures  were 
performed  on  the  algorithm  at  all  three  resolutions  (5,  30,  100  meters): 


212 


•  define  the  projection 

•  classify  the  data 

•  extract  the  slope  category  values  for  the  field  locations. 

These  slope  category  values  for  the  field  locations  were  then  used  in  the  statistical 
comparisons. 

The  ERDAS  IMAGINE  Method 

The  original  5-meter  point  file  generated  in  ARC/INFO  from  the  ASCII  output, 
was  imported  into  IMAGINE.  An  image  file  was  created  at  a  5-meter  cell  size  and 
the  resample  function  was  used  to  thin  the  data  set  to  30  and  100-meter  resolutions. 
Also  imported  was  the  field  location  point  file  generated  in  ARC/INFO.  The  same 
procedures  as  in  ARC/INFO  were  used  for  the  elevation  data  sets  at  all  three 
resolutions. 

Using  the  image  info  utility,  the  projection,  datum,  and  units  of  measure  were 
defined  for  the  elevation  data  sets  (as  well  as  the  field  location  data  set)  in  accordance 
with  those  parameters  used  during  the  extraction  process.  The  raster  recode  function 
was  used  to  reclassify  the  data  into  TTADB  slope  categories.  Finally,  a  spatial  model 
was  created  with  the  model  utility  using  the  conditional  function.  Using  the  slope 
data  and  field  points  as  input,  the  conditional  function  outputs  the  pixel  value  of  the 
first  input  grid  for  even-  pixel  value  existing  in  the  second  input  grid.  Therefore,  a 
slope  value  was  output  for  every  field  location.  The  pixel  to  table  function  was  used 
on  the  resulting  image  file  to  create  an  ASCII  file  that  contained  the  x,  y,  z  values  for 
each  pixel.  A  z  value  of  zero  indicated  that  the  pixel  was  not  a  field  location  point. 
To  extract  the  field  location  points,  the  UNIX  awk  command  was  used  to  output  the 
x,  y,  z  values  for  each  pixel  with  a  non-zero  z  value.  These  values  were  then  used  in 
the  statistical  comparisons. 

STATISTICAL  ANALYSIS 

The  slope  values  generated  by  the  algorithms  at  the  selected  points  were  compared 
with  the  values  obtained  in  the  field  by  using  a  confusion  matrix.  The  generated 
values  were  matched  against  the  known  (field  measurement)  values,  allowing  an 
assessment  of  slope  category  accuracy  and  overall  accuracy.  These  results  were  used 
to  judge  the  performance  of  the  algorithm. 


213 


Table  3.--SampIe  confusion  matrix. 


Generated  Values 


A 

B 

C 

D 

E 

F 

Row  Total 

A 

32 

11 

43 

B 

4 

38 

6 

48 

C 

1 

15 

28 

4 

1 

49 

D 

1 

1 

11 

15 

5 

33 

E 

1 

11 

14 

26 

F 

2 

2 

Column 
Total 

38 

65 

48 

30 

20 

0 

201 

The  bold  numbers  in  the  sample  above  are  the  number  of  points  correctly 
classified.  This  means  that  at  a  given  point,  the  generated  value  matched  the  field 
measurements.  In  the  sample,  32  points  had  the  correct  "A"  value.  Eleven  points 
were  classified  as  "B"  values  by  the  algorithm  while  they  were  actually  "A"  values 
according  to  field  measurements. 

To  calculate  the  percent  correct  in  any  given  category,  divide  the  total  correctly 
classified  for  the  category  by  the  row  total  (i.e.,  32/43  =  0.744  or  74.4%  correctly 
classified  as  "A"  slope).  To  calculate  the  overall  accuracy,  the  total  correctly 
classified  for  the  matrix  (sum  of  the  diagonals)  is  divided  by  the  total  of  all  the  values 
in  the  matrix  (i.e.,  (32  +  38  +  28  +  15  +  14)/201  =  0.632  or  63.2%  of  the  values  were 
correctly  classified  by  this  algorithm). 

RESULTS 

The  prediction  accuracies  of  the  methods  tested  are  shown  in  Table  4. 

DISCUSSION 

The  slope  generation  algorithm  embedded  in  ARC/INFO  outperformed  all  others 
at  5  and  100-meter  resolutions.  At  30-meter  resolution,  the  ARC/INFO  algorithm 
also  performed  well  though  was  slightly  less  accurate  than  the  WES  algorithm 
(74.6%  versus  76.1%).  The  WES  algorithm  was  a  consistently  good  performer,  but 
failed  to  classify  any  "B"  slopes  at  5  meters,  indicating  a  possible  flaw  in  the 
software.  Further  testing  must  be  done  before  any  recommendation  can  be  made. 
All  of  the  algorithms  tested,  except  the  WES  tended  to  underestimate  slope.  The 
WES  algorithm  overestimated  slope,  especially  in  the  "B"  category.  The  manually 


214 


Table  4. -Results  of  testing 


ELEVATION 
RESOLUTION 

METHOD 

CATEGORY  ACCURACY  (%) 

/  ^  %  /  1  '  11  All 

OVERALL 
ACCURACY 
(%) 

A 

B 

C 

D 

E 

F 

5  Meters 

ARC/INFO 

74.4 

79.2 

57.1 

45.5 

53.8 

0.0 

63.2 

IMAGINE 

65.8 

63.0 

41.2 

38.2 

48.0 

0.0 

50.8 

WES 

75.0 

0.0 

83.3 

45.7 

69.2 

0.0 

51.7 

1:24,000 

81.6 

46.7 

48.1 

37.5 

13.0 

60.0 

48.3 

TTADB 

84.6 

38.6 

32.1 

37.8 

40.0 

0.0 

44.8 

30  Meters 

ARC/INFO 

86.7 

85.2 

68.8 

56.2 

63.6 

-- 

74.6 

IMAGINE 

93.6 

82.0 

59.6 

48.4 

50.0 

— 

70.3 

WES 

79.5 

84.9 

71.4 

69.7 

68.2 

— 

76.1 

1:24,000 

90.9 

69.8 

58.6 

45.5 

12.0 

60.1 

TTADB 

88.6 

51.1 

41.8 

51.4 

54.2 

— 

57.1 

100  Meters 

ARC/INFO 

93.5 

83.9 

70.7 

46.4 

23.1 

73.1 

IMAGINE 

93.2 

75.4 

52.5 

41.4 

7.7 

63.4 

WES 

65.9 

76.9 

77.8 

66.7 

50.0 

71.6 

1:24,000 

93.3 

64.4 

56.9 

36.4 

13.3 

60.1 

TTADB 

88.9 

44.7 

39.7 

48.5 

66.7 

55.2 

compiled  overlays  also  overestimated  slope. 

Overall,  accuracy  of  the  algorithms  at  30  and  100-meter  resolutions  was  better  than 
the  accuracy  achieved  at  5  meters,  when  compared  with  the  field  collected  data.  A 
possible  reason  for  this  lies  in  the  accuracy  of  the  coordinates  obtained  using 
differential  GPS.  Although  the  method  produced  ground  coordinates  with  an 
accuracy  of  +/-  2.5  meters,  this  may  not  have  been  sufficient  accuracy  for 
comparisons  to  the  5-meter  grid.  Due  to  the  2.5-meter  uncertainty  in  the  location  of 
the  measuring  point,  exact  correlation  with  the  automated  slope  prediction  grid  was 
not  possible.  This  discrepancy  may  be  responsible  for  misclassification  of  some 
points. 

As  shown  in  Table  4,  the  manually  compiled  overlays  at  1 :24,000  and  1 :50,000 
scale  (TTADB)  were  less  accurate  slope  predictors  than  all  of  the  algorithms.  As 
expected,  the  larger  scale  overlay  (1 :24,000)  did  predict  better  than  the  smaller  scale 
overlay  (1:50,000).  Part  of  the  reason  for  the  observed  lower  accuracy  of  these 
overlays  is  the  minimum  compilation  size  of  each  polygon.   For  this  study,  the 


215 


smallest  area  retained  during  compilation  was  250  by  250  meters  on  the  ground. 
This  minimum  size  affects  the  prediction  accuracy  though  the  overriding  factor  is 
that  manual  compilation  is  clearly  less  accurate  than  automated  approaches. 

All  algorithms  tested  to  this  point  are  grid-based.  The  need  to  examine  algorithms 
utilizing  triangulated  irregular  networks  (TINs)  to  compute  slope  still  exists. 
Additionally,  TINs  using  geomorphic  data  (such  as  drainage  networks)  must  be 
tested  to  see  if  this  additional  data  improves  the  accuracy  of  the  slope  predictions. 

CONCLUSIONS 

1 .  Use  of  grid-based  slope  algorithms  alone  will  produce  point  slope  predictions 
that  match  the  TTADB  slope  categories  approximately  75%  of  the  time. 

2.  The  WES  Maximum  slope  and  the  ARC/INFO  algorithms  performed  similarly 
except  for  the  "B"  slope  anomaly  for  the  WES  Maximum  slope  algorithm  noted 
earlier  in  the  discussion. 

3.  The  ERDAS  IMAGINE  algorithm  produced  slightly  lower  prediction  accuracies 
than  the  WES  Maximum  slope  and  the  ARC/INFO  algorithms. 

4.  All  of  the  automated  slope  prediction  algorithms  outperformed  the  manual  slope 
compilation  approach. 

5.  Use  of  TIN-based  algorithms  with  and  without  geomorphic  data  may  increase 
prediction  accuracy. 

ACKNOWLEDGMENTS 

The  authors  of  this  paper  would  like  to  thank  Louis  Fatale  and  James  Ackeret  for 
their  excellent  field  support,  and  Patrick  Nguyen  for  his  system  support.  We  also 
greatly  appreciate  the  leadership  and  support  of  Mr.  Jeffrey  Messmore. 

REFERENCES 

Burrough.  P. A.  1986.  Principles  of  Geographical  Information  Systems  for  Land 
Resources  Assessment,  New  York,  NY:  Oxford  University  Press,  pp.  39-56. 

Environmental  Systems  Research  Institute.  1994.  Help  documentation  for 
ARC/INFO  7.0.2,  Redlands.  CA. 

ERDAS  Inc.  1995.  ERDAS  Field  Guide,  Atlanta,  GA.  pp.  322-324. 

Horn,  B.K.P.  1969.  Hill  shading  and  the  reflectance  map,  Proceedings  of  the  I.E.E.E. 
pp.  14-47. 

Skidmore.  A.K.  1989.  A  comparison  of  techniques  for  calculating  gradient  and 
aspect  from  a  gridded  digital  elevation  model,  International  Journal  of 
Geographical  Information  Systems,  Vol.  3,  No.  4.  pp.  323-334. 


216 


Experimental  Development  of  a  Model 
of  Vector  Data  Uncertainty 

G.J.  Hunter1,  B.  Hock2,  M.  Robey3  and  M.F.  Goodchild4 


Abstract. — In  a  recent  paper  by  Hunter  and  Goodchild  (1996),  a  model 
of  vector  data  uncertainty  was  proposed  and  its  conceptual  design  and 
likely  manner  of  implementation  were  discussed.  The  model  allows  for 
probabilistic  distortion  of  point,  line  and  polygon  features  through  the 
creation  of  separate,  random  horizontal  positional  error  fields  in  the  x  and 
y  directions.  These  are  overlaid  with  the  vector  data  so  as  to  apply 
coordinate  shifts  to  all  nodes  and  vertices  to  establish  new  versions  of  the 
original  data  set.  By  studying  the  variation  in  the  family  of  outputs 
derived  from  the  distorted  input  data,  an  assessment  may  be  made  of  the 
uncertainty  associated  with  the  final  product.  This  paper  is  a  contin- 
uation of  that  initial  work  and  discusses  the  experimental  development 
undertaken  thus  far  to  implement  the  model  in  practice. 


INTRODUCTION 

In  a  recent  paper  by  Hunter  and  Goodchild  (1996),  a  model  of  vector  data 
uncertainty  was  proposed  and  its  conceptual  design  and  likely  manner  of 
implementation  and  application  were  discussed.  This  paper  is  a  continuation  of 
that  initial  work  and  discusses  the  experimental  development  since  undertaken  to 
implement  the  model  in  practice.  In  the  context  of  this  research,  we  suggest  there 
is  a  clear  distinction  to  be  made  between  'error'  and  'uncertainty',  since  the 
former  implies  that  some  degree  of  knowledge  has  been  attained  about  differences 
between  actual  results  or  observations  and  the  truth  to  which  they  pertain.  On  the 
other  hand,  'uncertainty'  conveys  the  fact  that  it  is  the  lack  of  such  knowledge 
which  is  responsible  for  hesitancy  in  accepting  those  same  results  or  observations 
without  caution,  and  the  term  'error'  is  often  used  when  it  would  be  more 
appropriate  to  use  'uncertainty'. 

The  model  that  has  been  developed  can  be  defined  as  a  stochastic  process 
capable  of  generating  a  population  of  distorted  versions  of  the  same  reality  (such 
as  a  map),  with  each  version  being  a  sample  from  the  same  population.  The 
traditional  Gaussian  model  (where  the  mean  of  the  population  estimates  the  true 
value  and  the  standard  deviation  is  a  measure  of  variation  in  the  observations)  is 
one  attempt  at  describing  error,  but  it  is  global  in  nature  and  says  nothing  about 
local  variations  or  the  processes  by  which  error  may  have  accumulated. 

The  model  applied  here  is  viewed  as  an  advance  on  that  approach  since  it  has 
the  ability  to  show  spatial  variation  in  uncertainty,  and  the  capability  to  include  in 
its  realizations,  the  probable  effects  of  error  propagation  resulting  from  the 


Assistant  Professor,  Department  of  Geomatics,  The  University  of  Melbourne,  Victoria,  Australia 
Scientist,  Forest  Research  Institute,  Rotorua,  New  Zealand 

PhD  student,  Department  of  Geomatics,  The  University  of  Melbourne,  Victoria,  Australia 

Director,  National  Center  for  Geographic  Information  &  Analysis,  University  of  California,  Santa  Barbara,  CA 


217 


subsequent  algorithms  and  processes  that  have  been  applied  to  the  data.  This 
latter  point  is  particularly  important,  since  most  software  vendors  do  not  divulge 
the  algorithms  used  in  their  packages  for  commercial  reasons — which  prevents 
formal  mathematical  error  propagation  analysis  from  being  undertaken.  By 
studying  different  versions  of  the  products  created  by  the  model,  it  is  possible  to 
see  how  differences  in  output  are  affected  by  variations  in  input.  Elsewhere  in 
this  proceedings  a  grid-cell  version  of  the  model  (which  is  at  a  more  advanced 
stage  of  development),  has  been  described  and  applied  to  assess  the  uncertainty 
associated  with  landslide  susceptibility  indices  derived  from  a  simple  slope 
stability  model  (Murillo  and  Hunter,  1996). 

The  vector  data  model  involves  the  creation  of  two  independent,  normally 
distributed,  random  error  grids  in  the  x  and  y  directions.  When  combined,  these 
grids  provide  the  two  components  of  a  set  of  simulated  positional  error  vectors 
regularly  distributed  throughout  the  region  of  the  data  set  to  be  perturbed  (Figure 
1).  The  assumptions  made  by  the  authors  are  firstly  that  the  error  has  a  circular 
normal  distribution,  and  secondly  that  its  x  and  y  components  are  independent  of 
each  other.  The  grids  are  generated  with  a  mean  and  standard  deviation  equal  to 
the  data  producer's  estimate  for  positional  error  in  the  data  set  to  be  perturbed  (a 
prerequisite  for  application  of  the  model).  These  error  estimates,  for  example, 
might  come  from  the  residuals  at  control  points  reported  during  digitiser  setup,  or 
from  existing  data  quality  statements  such  as  those  that  now  accompany  many 
spatial  data  sets. 

By  overlaying  the  two  grids  with  the  vector  data  to  be  distorted,  x  and  y 
positional  shifts  can  be  applied  to  the  coordinates  of  each  node  and  vertex  in  the 
data  set  to  create  a  new,  but  equally  probable,  version  of  it.  Thus,  the 
probabilistic  coordinates  of  a  point  are  considered  to  be  (x  +  error,  y  +  error). 
With  the  distorted  version  of  the  data,  the  user  then  applies  the  same  set  of 
procedures  as  required  previously  to  create  the  final  product,  and  by  using  a 
number  of  these  distorted  data  sets  the  uncertainty  residing  in  the  end  product  is 
capable  of  being  assessed.  Alternatively,  several  different  data  sets  may  be 
separately  distorted  (each  on  the  basis  of  its  own  error  estimate)  prior  to  being 


Figure  1. — The  model  of  vector  data  uncertainty  uses  normally  distributed,  random 
error  grids  in  the  x  and  y  directions  to  produce  a  distorted  version  of  the  original  data. 


x  error  grid 


Distorted  data 


y  error  grid 


218 


combined  to  assess  final  output  uncertainty.  While  the  model  requires  an  initial 
error  estimate  for  creation  of  the  two  distortion  grids,  it  is  the  resultant  uncertainty 
arising  from  the  use  of  perturbed  data  due  to  simulation  which  is  under 
investigation  (in  conjunction  with  the  spatial  operations  that  are  subsequently 
applied) — hence  its  label  as  an  'uncertainty'  model. 

DEVELOPMENT  OF  THE  MODEL 

Choosing  the  Error  Grid  Spacing 

As  discussed  in  Hunter  and  Goodchild  (1996),  the  first  step  required  to 
implement  the  model  is  to  determine  an  appropriate  error  grid  spacing.  If  it  is  too 
large,  the  nodes  and  vertices  of  small  features  in  the  source  data  will  receive 
similar-sized  shifts  in  x  and  y  during  perturbation  and  the  process  will  not  be 
random,  giving  rise  to  unwanted  local  autocorrelation  between  shifts  (Figure  2). 
Conversely,  if  it  is  too  small  then  processing  time  can  increase  dramatically  as 
additional  grid  points  are  needlessly  processed. 


Figure  2. — If  the  error  grid  spacing  (d)  is 
too  large,  then  unwanted  local  autocorrelation 
between  shifts  may  occur  in  small  features. 


The  authors  suggest  an 
appropriate  spacing  be  selected  from 
one  of  the  following  options:  (1)  the 
standard  deviation  of  the  horizontal 
positional  error  for  the  source  data; 
(2)  a  distance  equal  to  0.5mm  at 
source  map  scale  where  the  data  has 
been  digitized;  or  (3)  a  threshold 
value  smaller  than  the  user  would  care 
to  consider  given  the  nature  of  the 
data  to  be  processed.  Thus  far,  the 
authors  have  found  the  standard 
deviation  (as  supplied  by  the  data 
producer)  to  be  a  useful  value  to 
apply,  since  it  tends  to  be  smaller  than 
(2)  and  therefore  more  conservative, 
and  of  similar  magnitude  to  practical 
estimates  of  (3). 


Generating  the  Initial  Error  Grids 

Using  the  development  diagram  on  the  following  page  as  a  guide  for  the 
remaining  discussion  (Figure  3),  the  next  step  is  to  generate  the  x  and  y  error 
grids.  The  Arc/Info  and  Arc  Grid  software  modules  were  used  as  the 
development  platform  and  the  commands  referred  to  in  the  paper  apply  to  these 
packages.  To  ensure  the  grids  completely  cover  the  extent  of  the  source  data,  the 
dimensions  of  the  grids  were  pre-determined  by  setting  a  window  equal  to  the 
data  set's  dimensions  (SETWINDOW  command),  and  the  cell  size  equivalent  to 
the  chosen  grid  spacing  (SETCELL  command).  Two  grids  (named  accordingly 
for  the  x  and  y  directions)  were  created  automatically  using  these  parameters  and 
then  populated  with  randomly  placed,  normally  distributed  values  having  a  mean 
of  zero  and  a  standard  deviation  defined  by  the  data  producer  (NORMAL 


219 


Figure  3. — Development  diagram  for  the  uncertainty  model. 


command).  It  should  be  noted  here  that  selection  of  the  standard  deviation  for  the 
grid  spacing  does  not  affect  the  random  population  of  the  grids.  The  two  grids  are 
only  temporary  and  will  require  further  refinement  before  being  used  to  perturb 
the  vector  data. 

Reduction  of  Error  Grid  Size 

As  a  means  of  minimizing  processing  time  the  number  of  cells  in  the  error 
grids  needs  to  be  reduced,  since  unless  the  vector  data  set  is  extremely  dense  there 
will  be  many  cells  that  are  not  required  during  operation.  To  achieve  this,  the 
original  vector  data  are  converted  to  grid  format  (POINTGRID  or  LINEGRID 
commands)  to  form  a  temporary  masking  grid  which  only  contains  'live'  cells  that 
the  source  data  either  lie  within  or  pass  through.  Note  that  polygons  are  also 
processed  by  LINEGRID  since  only  boundary  strings  need  be  perturbed.  The  cell 
attribute  that  has  been  passed  during  rasterization  is  unimportant  since  the  grid 
will  only  be  used  for  masking  purposes.  All  other  (non-contributing)  cells  are 
deliberately  given  a  NOD  ATA  value. 

However,  a  problem  arises  here  with  using  a  masking  grid  that  contains  only 
single  cells  (representing  point  data)  or  1 -cell-wide  strings  (representing  lines  or 
polygons).  As  mentioned  in  Hunter  and  Goodchild  (1996),  there  is  the  possibility 
that  the  magnitude  and  direction  of  adjacent  x  and  y  error  grid  shifts  may  cause 
them  to  overlap  each  other,  resulting  in  possible  loss  of  topological  integrity  in  the 
original  data  when  applied  to  it  (such  as  the  transposition  of  adjacent  vertices 


220 


causing  unwanted  loops). 

The  solution  requires  filtering  of  any  such  'offending'  pairs  of  shifts  which 
requires  that  the  masking  grid  be  extended  for  a  distance  of  at  least  five  standard 
deviations  either  side  of  the  original  data — given  that  with  the  normal  distribution 
it  is  rare  for  a  value  to  occur  greater  than  this  distance  from  the  mean.  This 
buffering  of  the  masking  grid  will  permit  neighboring  error  grid  shifts  to  be  used 
in  a  subsequent  filtering  process. 

To  achieve  this,  the  grid  spacing  is  compared  with  the  standard  deviation  of 
the  data  and  the  equivalent  number  of  cells  is  calculated  (remembering  that  the 
grid  spacing  will  not  always  equal  the  standard  deviation).  For  example,  using  a 
standard  deviation  of  20  m  and  a  grid  spacing  of  30  m,  the  correct  number  of  cells 
equals  1(100/30)  +  11  =  4  (with  the  extra  cell  and  the  modulus  sign  ensuring 
'rounding  up'  of  the  answer).  Using  this  value,  a  spread  function  is  applied  to  the 
masking  grid  (EUCDISTANCE  command).  A  new  masking  grid  is  created 
during  the  process  and  any  (previously)  NOD  ATA  cells  affected  by  the  operation 
are  automatically  returned  to  active  status. 

The  masking  grid  is  then  overlaid  with  the  initial  error  grids  to  provide 
reduced  versions  of  the  x  and  y  grids  that  contain  only  shift  values  lying  within  the 
required  regions.  Finally,  the  error  grids  are  expanded  by  the  width  of  a  cell  on 
all  sides  (and  given  NOD  ATA  values)  to  support  later  processing  (SELECTBOX 
command). 

Filtering  the  Error  Grids 

To  preserve  topological  integrity  in  the  source  data  upon  distortion,  some 
means  of  adjustment  must  be  introduced  to  control  the  magnitude  of  positional 
shifts  between  neighbouring  points  in  the  x  and  y  error  grids.  As  previously 
mentioned,  if  this  does  not  occur  there  is  a  likelihood  that  neighbouring  points  in 
the  original  data  set  may  be  transposed  in  position — causing  unwanted  'fractures' 


Figure  4. — In  (a),  uncontrolled  shifts  between  neighbouring  error  grid  points  can  cause 
unwanted  'fractures'  or  transposition  of  features.  In  (b),  'fractures'  occur  when  the 
difference  between  neighbouring  shifts  is  larger  than  their  separation  (d).  In  (c),  a 
'fracture'  is  circled,  requiring  filtering  on  the  basis  of  neighbouring  shifts. 


221 


in  the  feature  structure  (Figure  4a).  Intuitively,  these  modifications  should  not 
occur  and  are  considered  unacceptable.  Originally,  the  authors  considered  using 
spatial  autocorrelation  to  constrain  the  jc  and  y  shifts,  but  this  would  have  had  the 
effect  of  unnecessarily  altering  all  shifts  in  the  error  grids  when  only  a  relatively 
small  number  needed  to  be  adjusted  in  practice.  Accordingly,  a  localised  filter 
was  seen  as  a  better  solution. 


Therefore,  a  routine  was  developed  to  test  the  difference  between  consecutive 
cells  (in  horizontal  or  row  sequence  for  the  jc  grid,  and  vertical  or  column 
sequence  for  the  y  grid)  to  determine  whether  the  absolute  value  of  the  difference 
between  them  was  greater  than  their  separation  distance  (Figure  4b).  If  so,  then  a 
'fracture'  is  possible  at  that  location  and  a  filter  must  be  applied  to  average  out  the 
shift  values  on  the  basis  of  their  neighbors.  The  procedure  is  iterative  and 
proceeds  until  no  'fractures'  exist  in  the  error  grids. 

For  example,  given  consecutive  shifts  Axo,  Axx,  Ax2  and  Ax3  in  any  row  of  the  x 
error  grid  (and  testing  the  difference  between  the  two  middle  shifts),  there  is 
potential  for  a  'fracture'  to  occur  if  Axx  -  Ax2  >d  (where  d  is  the  grid  spacing).  The 
adjusted  values  of  Axl  and  Ax2  are  computed  by  equation  (1): 

^■=A*.+Ax1+Ax2    and  ^2  =  A*1+A*2+As3  (J) 

The  justification  for  the  filtering  is  that  although  the  shifts  are  normally 
distributed,  the  choice  of  that  distribution  is  only  an  assumption  and  we  should 
accept  there  may  be  faults  in  its  selection  as  a  theoretical  model.  For  example, 
statisticians  developed  the  human  Intelligent  Quotient  (IQ)  using  a  normal 
distribution  with  a  mean  of  100  and  standard  deviation  of  20.  Yet  the  distribution 
is  deliberately  truncated  at  zero  to  disallow  negative  values  which  intuitively 
should  not  occur — even  though  they  will  be  predicted  by  the  model.  In  a  similar 
manner,  the  filtering  of  neighboring  error  shifts  to  avoid  'fractures'  is  simply  a 
truncation  of  the  distribution  to  cater  for  a  minority  of  outlying  values. 

The  filtering  was  implemented  by  developing  a  routine  that  moves  progress- 
ively through  each  error  grid  and  tests  the  difference  between  each  pair  of  cell 
values  (DOCELL  command).  If  the  condition  for  a  'fracture'  is  not  met,  the 
values  are  written  to  an  update  grid  and  the  next  pair  is  tested.  If  the  condition  is 
met,  the  filter  is  applied  and  new  values  are  computed  and  passed  to  the  update 
grid.  In  practice,  it  has  been  found  that  shifts  of  several  standard  deviations  in 
size  can  cause  several  iterations  of  the  process  to  be  required.  However,  the 
procedure  still  converges  fairly  rapidly. 

An  early  problem  encountered  was  the  situation  where  a  large  shift  existed  in 
a  perimeter  row  or  column  of  the  error  grid — requiring  an  extra  cell  value  (which 
did  not  exist)  to  be  input  to  the  filter  equation.  It  was  found  that  without  the 
placement  of  an  extra  row  or  column  on  each  boundary  of  the  grid,  the  absence  of 
a  cell  caused  the  filter  process  to  substitute  a  NOD  ATA  value  for  the  cell  to  be 
adjusted.  This  was  unacceptable  and  the  placement  of  an  extra  perimeter  cell 
around  the  error  grid  (with  NOD  ATA  values)  ensured  that  the  original  large  shift 
was  retained  unaltered.  This  was  considered  preferable  to  it  being  discarded. 
While  this  implies  that  some  nodes  and  vertices  may  receive  shifts  to  new 


222 


Figure  5. — The  x  andy  shifts  for  a  data  point  not  coinciding  with 
the  error  grid  are  calculated  using  bilinear  interpolation  based  on 
the  four  surrounding  grid  values. 

positions  outside  the  original  extent  of  the  data  set,  this  decision  remains 
consistent  with  the  original  concept  of  distorting  their  values  to  form  alternative, 
but  equally  probable  versions  of  the  data.  At  the  time  of  writing,  the  experimental 
development  of  the  model  has  not  proceeded  beyond  this  stage  and  the  following 
discussion  remains  conceptual. 

Calculation  of  Data  Point  Shifts 

In  the  final  step  of  the  model's  development,  values  in  the  error  grids  must  be 
transferred  to  the  data  set  being  perturbed.  Inevitably,  it  is  expected  that  few  if 
any  nodes  or  vertices  in  the  source  data  will  coincide  exactly  with  the  error  grid 
points,  and  a  method  is  required  for  calculating  x  and  y  shifts  based  on  the 
neighboring  values  in  the  grid.  A  simple  bilinear  interpolation  procedure  (Watson, 
1992)  is  proposed  in  which  the  x  and  y  shifts  assigned  to  each  point  are  calculated 
on  the  basis  of  the  respective  shifts  of  the  four  surrounding  grid  points  (Figure  5). 

An  ASCII  feature  file  containing  the  identifier  and  coordinates  of  each  data 
point  can  be  automatically  derived  (UNGENERATE  command),  and  the  four 
surrounding  error  shifts  will  be  determined  for  each  point  and  used  to  interpolate 
the  correct  values  to  be  applied.  The  distorted  coordinates  are  then  written  to  a 
new  file  and  when  all  records  are  processed,  the  file  topology  is  rebuilt.  Clearly, 
the  need  may  arise  for  high  performance  computers  to  be  employed  for  this  task, 
particularly  when  many  thousands  of  nodes  and  vertices  may  have  to  be 
perturbed,  and  current  research  into  GIS  and  supercomputers  (such  as  described  in 
Armstong,  1994)  is  being  investigated  by  the  authors. 

CONCLUSIONS 

This  paper  has  described  the  experimental  development  of  an  uncertainty 
model  for  vector  data,  which  operates  by  taking  an  input  data  set  of  point,  line  or 
polygon  features  and  then  applying  simulated  positional  error  shifts  in  the  x  and  y 
directions  to  calculate  new  coordinates  for  each  node  and  vertex.  In  effect  this 
produces  a  distorted,  but  equally  probable,  representation  of  the  data  set  that  can 
be  used  to  create  a  family  of  alternative  outputs,  usually  in  map  form. 


223 


Assessment  of  the  variation  in  the  outputs  can  used  to  provide  an  estimate  of  the 
uncertainty  residing  in  them,  based  on  the  error  in  the  source  data  and  its 
propagation  through  the  subsequent  algorithms  and  processes  employed. 

At  this  stage  of  the  research,  software  has  been  written  that  automates  the 
creation  of  the  error  grids  and  filters  their  values  to  avoid  conditions  that  would 
cause  loss  of  topological  integrity  in  the  original  data.  The  software  also  masks 
the  input  data  to  ensure  redundant  values  in  the  error  grids  are  not  unnecessarily 
processed.  Optimisation  of  the  process  is  still  required  to  achieve  faster 
operation.  The  next  stage  of  the  project  involves  development  of  an  algorithm 
which  enables  each  node  and  vertex  in  the  input  data  to  be  allocated  the  correct  x 
and  y  shift  from  the  error  grids,  which  yields  the  perturbed  version  of  the  source 
data.  The  final  stage  entails  testing  of  the  working  model,  refinement  of  its 
operation,  and  assessment  of  its  strengths  and  limitations. 

ACKNOWLEDGMENTS 

The  authors  wish  to  acknowledge  the  support  given  to  this  project  by  the  New 
Zealand  Forest  Research  Institute,  the  Australian  Research  Council,  and  the 
Australian  Federal  Department  of  Industry,  Science  &  Technology  (under  the 
Bilateral  Science  &  Technology  Collaboration  Scheme).  The  National  Center  for 
Geographic  Information  &  Analysis  is  supported  by  the  National  Science 
Foundation,  grant  SBR  88-10917,  and  this  research  constitutes  part  of  the 
Center's  Research  Initiative  7  on  Visualisation  of  Spatial  Data  Quality. 

REFERENCES 

Armstrong,  M.P.  1994.  "GIS  and  High  Performance  Computing".  Proceedings  of 

the  GIS/LIS  94  Conference,  Phoenix,  Arizona,  pp.  4-13. 
Hunter,  G.J  and  Goodchild,  M.F.  1996.  "A  New  Model  for  Handling  Vector  Data 

Uncertainty  in  GIS".  Urban  and  Regional  Information  Systems  Journal,  vol. 

8,  no.  1,  12  pp. 

Murillo,  M.  and  Hunter,  G.J.  1996.  "Modeling  Slope  Stability  Uncertainty:  A 
Case  Study  at  the  H.J.  Andrews  Long-Term  Ecological  Research  Site, 
Oregon".  Proceedings  of  the  2nd  International  Symposium  on  Spatial 
Accuracy  Assessment  in  Natural  Resources  and  Environmental  Modeling,  Fort 
Collins,  Colorado,  10  pp. 

Watson,  D.F.,  1992.  Contouring:  A  Guide  to  the  Analysis  and  Dispaly  of  Spatial 
Data  (Pergamon/Elsevier  Science:  New  York). 


224 


A  User-friendly  Tool  for  Error 
Modelling  and  Error  Propagation 
in  a  GIS  Environment 

Frank  Forier1  and  Frank  Canters2 


Abstract.  -  In  spite  of  a  growing  concern  for  spatial  database 
accuracy  today  most  commercial  GIS  software  products  do  not 
incorporate  tools  for  the  modelling  of  error  in  individual  data  layers 
and  for  the  tracking  of  error  when  data  layers  are  manipulated  and 
combined  in  GIS-based  spatial  analysis.  This  paper  reports  on  design 
considerations  that  have  been  taken  into  account  in  the  development 
of  a  system-independent  tool  for  error  modelling  and  error 
propagation  in  raster-formatted  categorical  data.  The  tool  aims  at  an 
optimal  use  of  available  methods  for  formal  error  propagation  and 
simulation  modelling.  Special  attention  is  given  to  the 
parameterization  of  different  error  models,  i.e.  how  to  make  optimal 
use  of  existing  information  on  error  in  the  source  data.  When 
simulating  error  in  the  data  layers  spatial  autocorrelation  in  the  error 
field  can  be  taken  into  account.  This  allows  a  more  realistic 
assessment  of  uncertainty  for  spatial  analysis  which  involves  the 
application  of  neighbourhood  or  global  operations.  

INTRODUCTION 

The  widespread  availability  and  use  of  geographical  information  systems  in  recent 
years  has  greatly  enhanced  the  opportunities  for  a  more  rationalized  approach  to 
spatial  decision  making.  There  is,  however,  a  growing  awareness  that  a  successful 
use  of  GIS  can  only  be  achieved  if  it  becomes  possible  to  attach  a  quality  label  to 
the  output  of  each  GIS  analysis.  Therefore  issues  of  spatial  data  quality,  error 
modelling  and  error  propagation  have  gained  considerable  interest  from  the  GIS 
research  community.  The  number  of  publications  on  the  subject  is  rapidly 
increasing  and  several  techniques  for  the  modelling  of  uncertainty  in  spatial  data 
sets  as  well  as  formal  methods  describing  the  mechanisms  of  error  propagation 
have  been  developed. 

Despite  the  impressive  volume  of  research  on  the  topic  of  error  in  spatial 
databases  there  is  still  a  lack  of  proper  tools  for  error  handling  in  commercial  GIS 
software.  The  aim  of  the  work  described  in  this  paper  is  to  bundle  existing 
techniques  for  error  modelling  and  error  propagation  in  a  software-independent 
tool.  In  designing  the  tool  the  following  aims  as  stated  by  Heuvelink  (1993)  are 
pursued:  1.  the  error  tool  should  not  replace  the  GIS  but  only  add  functionality;  2. 
the  error  tool  should  be  able  to  handle  a  variety  of  GIS  operations;  3.  the  error  tool 
must  have  a  user-friendly  interface  to  encourage  its  use;  4.  the  error  tool  should 
employ  efficient  routines  to  minimize  the  numerical  load. 


*  Centre  for  Cartography  and  GIS,  Department  of  Geography,  Vrije  Universiteit  Brussel,  Brussel,  Belgium. 
E-mail:  fforier@vnet3.vub.ac.be,  fcanters@vnet3.vub.ac.be 


225 


The  process  of  handling  uncertainty  in  spatial  data  involves  three  steps.  The  first 
step  consists  in  identifying  for  each  data  layer  what  kind  of  error  information  is 
available  or  can  be  extracted  from  other  data  sources.  The  type  of  error  information 
will  determine  which  assumptions  have  to  be  made  when  error  is  modelled  (Lanter 
and  Veregin  1992).  For  example,  if  thematic  error  is  merely  described  by  a  global 
index  of  classification  accuracy,  e.g.  the  PCC  index  or  kappa  statistic,  no  thematic 
or  spatial  differentiation  of  error  is  possible.  This  means  that  one  is  forced  to 
assume  that  error  levels  do  not  vary  from  class  to  class  and  are  uniformly 
distributed  over  space,  even  although  this  will  probably  not  be  the  case. 

The  second  step  in  the  error  handling  process  involves  the  definition  of  a 
conceptual  model  of  error.  The  choice  which  is  made  at  this  point  will  depend  to  a 
large  extent  on  the  kind  of  error  information  that  is  available  (step  1).  With  respect 
to  the  type  of  error  information  used  in  the  model  a  distinction  can  be  made  between 
global,  thematic  and  spatial/thematic  error  models.  The  choice  of  an  appropriate 
error  model  will  also  depend  on  the  error  propagation  technique  that  will  be  applied 
(step  3).  In  this  connection  a  distinction  should  be  made  between  purely  descriptive 
and  stochastic  error  models. 

Finally  an  appropriate  error  propagation  function  has  to  be  selected  to  verify  how 
error  is  propagated  through  the  GIS  operation.  In  the  tool  which  is  described  in  this 
paper  two  techniques  for  error  propagation  are  considered,  i.e.  formal  error 
propagation  and  simulation  modelling.  Formal  error  propagation  uses  explicit 
mathematical  models  which  describe  the  mechanisms  of  error  propagation  for 
particular  GIS  operations  (Veregin  1994).  The  formal  approach  has  the  advantage 
of  easy  applicability.  Starting  from  a  number  of  error  indices  which  describe  the 
quality  of  the  data  layer(s)  involved  in  the  operation  a  mathematical  function  is 
applied  which  directly  calculates  one  or  more  error  indices  for  the  output  layer.  The 
problem  with  formal  error  propagation  is  that  very  little  is  actually  known  about 
error  propagation  mechanisms.  Hence  error  propagation  functions  are  only 
available  for  a  small  number  of  relatively  simple  GIS  operations  and  prove  to  be 
rather  restrictive  with  respect  to  the  type  of  error  models  that  can  be  used.  If  formal 
error  propagation  cannot  be  applied  simulation  modelling  is  the  alternative.  The 
main  advantage  of  simulation  modelling  is  its  overall  applicability  regardless  of  the 
type  of  GIS  operator  which  is  involved.  A  disadvantage,  however,  is  the  high 
numerical  load  which  makes  that  simulation  modelling  is  less  practical  in  connection 
with  real-time  error  propagation. 


IDENTIFICATION  AND  EXTRACTION  OF  ERROR  INFORMATION 

One  of  the  most  important  obstacles  when  handling  uncertainty  in  GIS  is  the  lack 
of  knowledge  about  the  error  which  is  present  in  the  source  data.  The  type  of  error 
information  which  is  available  will  define  what  kind  of  error  modelling  can  be 
applied,  whether  error  can  be  differentiated  thematically  and/or  spatially,  and  as 
such  will  largely  determine  the  quality  of  the  whole  error  modelling  process. 
Because  the  problem  of  error  propagation  in  GIS  has  long  been  neglected 
information  on  the  quality  of  the  data  is  usually  very  sparse.  Ideally,  for  each 
source  layer  in  a  GIS  a  map  showing  the  spatial  distribution  of  error  should  be 
available.  Yet  very  often  only  global  indicators  of  data  quality  are  available 
providing  no  indication  on  local  variations  in  the  error  field. 

For  categorical  data  information  on  error  is  usually  described  by  a  confusion 
matrix  or  classification  error  matrix.  The  matrix  is  constructed  by  cross-tabulation 


226 


of  encoded  and  actual  attribute  values  for  a  sample  of  locations.  The  element  cy  in 
the  matrix  represents  the  number  of  cells  assigned  to  class  i  that  actually  belong  to 
class  j.  A  range  of  measures  of  accuracy  such  as  the  PCC  (Proportion  Correctly 
Classified)  index  and  the  kappa  statistic  (both  global  indices  of  classification 
accuracy),  and  the  user's  and  producer's  accuracy  for  each  class  can  be  deduced 
from  the  matrix  (Congalton  1991).  The  confusion  matrix  reports  error  on  the  level 
of  different  thematic  classes.  Its  major  drawback  is  that  it  provides  no  information 
on  within-class  variations  in  the  error  field.  The  confusion  matrix  will  therefore 
only  support  the  application  of  thematically  differentiated  error  models.  Spatially 
differentiated  models  require  extra  information  which  is  not  contained  in  the  matrix. 

Categorical  maps  are  usually  not  accompanied  by  a  description  of  error  in  the 
form  of  a  confusion  matrix.  When  attempting  to  model  error  in  GIS  source  layers 
that  are  derived  from  existing  maps  it  is  therefore  often  necessary  to  extract  error 
information  from  a  variety  of  other  sources.  Sometimes  information  on  error  can  be 
found  in  reports  which  are  delivered  with  map  documents.  Fisher  (1991),  for 
example,  showed  how  information  on  soil  map-unit  inclusions  from  soil  survey 
reports  can  be  used  to  build  a  confusion  matrix.  Information  on  uncertainty  can  also 
be  derived  from  field  inventories  (Aspinall  and  Pearson  1994)  or  from  more 
detailed  maps.  Overlaying  the  original  map  with  a  more  detailed  spatial  reference 
not  only  allows  to  derive  a  confusion  matrix,  but  also  makes  it  possible  to  obtain 
useful  information  on  the  spatial  structure  of  uncertainty,  for  example,  by 
calculating  the  mean  size  of  inclusions  of  each  class  in  the  reference  map  within 
each  class  in  the  original  source  map.  This  kind  of  information  is  of  utmost 
importance  for  the  application  of  spatially  differentiated  error  models,  as  will  be 
shown  later.  Interesting  information  may  also  be  provided  by  related  spatial  data. 
Fisher  (1989)  demonstrated  how  expert  knowledge  about  the  relation  between  soil 
and  other  physical  data  can  be  used  to  predict  the  spatial  distribution  of  error  in  soil 
maps  from  the  mapped  distribution  of  these  related  variables. 

In  developing  the  tool  which  is  referred  to  in  this  paper  special  effort  has  been 
made  to  provide  the  user  with  a  set  of  functions  which  allow  to  derive  error 
information  for  categorical  maps  from  more  detailed  or  related  map  documents.  If 
no  information  on  error  is  available  or  can  be  derived  from  other  sources  the  user  is 
offered  the  opportunity  to  introduce  uncertainty  in  the  data  layers.  This  allows  to 
carry  out  sensitivity  analysis  enabling  the  user  to  examine  to  what  extent  the  output 
of  a  sequence  of  GIS  operations  is  sensitive  to  specified  levels  of  uncertainty  in  the 
source  data. 


CHOOSING  AN  APPROPRIATE  ERROR  MODEL 

Depending  on  the  error  information  which  is  available  three  types  of  error 
modelling  can  be  applied:  global,  thematic  and  spatial/thematic  modelling.  For 
global  and  thematic  modelling  the  parameterization  of  the  error  models  is  based  on 
well-known  error  indices  that  can  be  derived  from  the  confusion  matrix.  Some  error 
propagation  techniques  use  the  entire  confusion  matrix  as  a  model  of  error  (Veregin 
1995).  Spatial/thematic  modelling  requires  additional  information  on  the  distribution 
of  error  within  each  thematic  class. 

In  global  error  modelling  thematic  error  is  described  by  one  single-valued  index 
which  is  used  as  a  model  parameter.  The  best  known  error  index  that  can  be  derived 
from  the  confusion  matrix  is  the  PCC  (Proportion  Correctly  Classified)  index 
which  represents  the  probability  that  a  location  selected  at  random  is  correctly 


227 


classified.  The  PCC  index  is  defined  as  the  sum  of  the  diagonal  elements  in  the 
confusion  matrix  divided  by  the  total  number  of  samples.  It  has  a  value  between  0 
and  1,  corresponding  with  maximum  and  no  error  respectively.  Alternatively  the 
kappa  statistic  may  be  used  (Congalton  1991).  The  kappa  index  allows  to  account 
for  change  agreement  and  will  be  equal  to  zero  if  agreement  is  equal  to  what  is 
expected  in  a  random  case.  If  agreement  is  less  it  will  be  negative,  if  agreement  is 
perfect  it  will  be  equal  to  1.  The  PCC  index  and  the  kappa  statistic  provide  an  over- 
all measure  of  error.  They  do  not  describe  how  error  may  vary  from  class  to  class. 
Hence  in  global  error  modelling  one  is  obliged  to  assume  that  error  levels  are  the 
same  for  all  classes  and  that  confusion  is  equal  among  each  pair  of  classes.  If  real 
error  conditions  largely  deviate  from  these  assumptions  error  propagation  results 
based  on  global  error  modelling  may  be  of  little  use. 

Thematic  error  modelling  accounts  for  variations  in  the  level  and  type  of  error  for 
different  classes.  Depending  on  the  error  propagation  technique  which  will  be  used 
specific  information  on  thematic  error  for  different  classes  will  be  derived  from  the 
confusion  matrix.  For  some  GIS  operations  formal  error  propagation  functions 
have  been  defined  which  propagate  the  entire  confusion  matrix  of  the  source 
layer(s)  through  the  operation  to  produce  a  confusion  matrix  for  the  output  layer 
(Veregin  1995).  This  means  that  all  error  information  which  is  present  in  the 
confusion  matrix  is  actually  used  in  the  modelling.  Hence  the  confusion  matrix 
itself  can  be  considered  as  the  error  model.  Unfortunately  at  present  formal  error 
propagation  functions  are  only  available  for  a  small  set  of  simple  GIS  operations.  If 
other  operations  are  involved  simulation  modelling  is  the  only  alternative. 

Simulation  modelling  requires  the  definition  of  a  stochastic  error  model  which 
allows  to  produce  perturbed  versions  of  the  original  data  layer  using  available  error 
information.  Stochastic  error  models  for  categorical  data  randomly  assign  each  cell 
in  the  data  layer  to  one  of  the  classes  according  to  a  priori  defined  class  membership 
probabilities  (Goodchild  and  Wang  1989,  Fisher  1991).  For  all  cells  which  have 
been  assigned  to  a  given  class  in  the  original  source  data  membership  probabilities 
for  all  classes  can  be  derived  from  the  confusion  matrix  by  dividing  the  elements  in 
the  corresponding  row  of  the  matrix  by  the  total  number  of  sample  locations  that 
have  originally  been  assigned  to  the  class.  Using  the  obtained  class  membership 
probabilities  the  most  obvious  way  of  introducing  uncertainty  into  the  source  layer 
is  to  assign  classes  independently  in  each  cell.  As  class  membership  probabilities 
for  a  cell  always  add  to  one  each  class  can  be  matched  to  a  sub-interval  in  the  range 
0-1  with  its  length  proportional  to  the  probability  of  that  class.  For  each  cell  a 
random  number  in  the  range  0-1  is  drawn  and  the  cell  is  assigned  to  the  class  that 
corresponds  with  the  interval  including  this  random  number.  Applying  a  sequence 
of  GIS  operations  to  the  perturbed  data  layer  will  yield  an  outcome  that  differs  from 
the  result  obtained  with  the  original  source  data.  Differences  in  outcome  which  are 
obtained  by  repeating  this  process  a  large  number  of  times  will  indicate  how 
uncertainty  is  propagated  through  the  analysis. 

Whether  formal  error  propagation  techniques  or  simulation  modelling  is  applied, 
the  use  of  an  error  model  which  is  solely  based  on  class- specific  information 
derived  from  the  confusion  matrix  ignores  the  possible  presence  of  autocorrelation 
in  the  error  field.  In  fact  a  model  which  only  permits  thematic  differentiation  of 
error  implicitly  assumes  that  error  distribution  is  purely  random,  in  other  words, 
that  cells  which  are  assigned  to  class  i  but  actually  belong  to  class  j  do  not  tend  to 
cluster  in  space.  In  most  cases  this  assumption  will  not  be  valid.  For  categorical 
data  only  one  formal  error  propagation  method  is  known  to  these  authors  which 
accounts  for  the  presence  of  autocorrelation  in  the  error  field.  The  method  allows  to 


228 


estimate  the  accuracy  of  a  buffer  operation  from  a  number  of  characteristics 
describing  the  source  data  and  the  error  which  is  associated  with  it  (Veregin  1994). 
Goodchild  et  al.  (1992)  presented  a  stochastic  error  model  for  categorical  data 
which  allows  to  control  the  level  of  spatial  dependence  by  inducing  correlation 
between  outcomes  in  neighbouring  cells.  The  mechanism  of  class  assignment  is 
again  based  on  the  cell's  class  membership  probabilities.  Yet  instead  of  drawing  a 
random  number  in  the  range  0-1  for  each  cell  independently  a  spatially  correlated 
random  field  is  generated  by  means  of  a  first  order  autoregressive  model  defined  on 
a  regular  grid: 

Zij  =  q  (zi-i,j  +  Zi+ij  +  zij.i  +  zij+i)  +eij  (1) 

where  zy  is  the  value  of  the  random  error  field  at  position  i,j,  q  (<  0.25)  is  a 
parameter  that  indicates  the  level  of  autocorrelation  and  where  ey  is  a  spatially 
independent  Gaussian  noise  field  with  mean  0  and  standard  deviation  unity. 
Heuvelink  (1992)  presented  a  simple  and  robust  iterative  method  for  the  generation 
of  random  fields  of  type  (1)  which  avoids  the  computational  overhead  associated 
with  the  traditional  inversion  method  (Haining  et  al.  1983).  Since  the  marginal 
distribution  of  the  correlated  random  field  is  of  the  Gaussian  type  with  zero  mean 
and  a  standard  deviation  which  is  function  of  the  parameter  q  a  simple 
transformation  allows  to  obtain  a  uniform  distribution  in  the  range  0-1.  This 
guarantees  that  class  assignment  for  each  cell  is  proportional  to  the  a  priori  defined 
class  membership  probabilities. 

Although  it  takes  spatial  correlation  into  account  Goodchild's  model  has  two 
important  drawbacks.  First  of  all,  it  is  only  applicable  to  two  classes.  If  the  model 
is  used  on  more  than  two  classes  the  inclusions  which  are  simulated  in  the  original 
map  units  will  be  spatially  nested  (figure  la).  Secondly  there  is  no  simple  way  of 
defining  a  proper  value  for  the  parameter  q.  Although  the  average  size  of  patches  in 
the  simulated  layer  is  clearly  related  to  the  value  of  q  it  is  not  possible  to  define  a 
one-to-one  correspondence  between  both  as  class  membership  probabilities  disturb 
the  relationship.  Based  on  Goodchild's  ideas  a  method  is  proposed  which  allows  to 
simulate  uncertainty  in  categorical  maps  with  more  than  two  classes.  For  each  class 
in  the  source  map  the  method  takes  into  account  the  type  and  the  mean  size  of 
inclusions  as  well  as  their  proportion  within  the  source  map  units.  Instead  of 
simulating  inclusions  of  different  classes  simultaneously  using  one  random  field 
inclusions  are  generated  sequentially  (one  class  at  a  time),  each  time  using  a 
different  random  field.  The  advantage  of  sequential  simulation  is  that  adjacent 
inclusions  will  occur  for  each  pair  of  classes,  independent  of  the  sequence  of 
simulation  (figure  lb). 

The  use  of  different  random  fields  also  allows  to  control  the  level  of 
autocorrelation  for  each  category  and  therefore  provides  a  mechanism  to  adjust  the 
size  of  inclusions.  Unfortunately,  as  has  been  mentioned  already,  it  is  very  difficult 
to  define  a  clear  relationship  between  the  level  of  autocorrelation  in  the  random  field 
and  the  mean  size  of  simulated  inclusions.  When  sequential  simulation  is  applied 
the  relationship  will  not  only  be  influenced  by  the  class  membership  probabilities, 
yet  also  by  the  exact  sequence  of  class  simulation.  As  the  level  of  autocorrelation 
which  is  required  to  generate  inclusions  of  a  given  size  cannot  be  defined  a  priori, 
an  iterative  method  is  used.  Starting  with  a  random  field  with  a  low  level  of 
correlation  inclusions  of  one  class  are  simulated.  If  the  mean  size  of  the  inclusions 
is  smaller  than  the  proposed  size  the  process  is  repeated  with  a  random  field  which 
has  a  higher  level  of  correlation,  until  the  proposed  size  of  inclusions  is  reached. 


229 


The  major  drawback  of  this  procedure  is  its  computational  load  since  dealing  with 
one  class  already  involves  the  use  of  different  correlated  random  fields.  On-line 
generation  of  these  fields  would  be  impossible.  That  is  why  a  large  set  of  correlated 
random  fields  was  generated  for  different  levels  of  autocorrelation  and  permanently 
stored  on  disk.  During  the  simulation  process  random  fields  with  a  required  level  of 
correlation  are  randomly  picked  from  this  set.  At  present  the  prototype  tool  for  error 
modelling  which  has  been  implemented  uses  15  sets  of  50  correlated  random  fields, 
each  set  corresponding  to  a  different  level  of  autocorrelation.  The  fact  that  only  50 
random  fields  are  defined  for  each  level  has  to  do  with  limitations  in  storage 
capacity.  The  size  of  the  fields  was  chosen  large  enough  (1500  x  1500)  to  allow  for 
their  use  in  a  large  variety  of  applications. 


□    A     ■    B     ■  C 


Figure  1.-  Simulation  of  classes  B  and  C  into  class  A  with  probabilities  pA  =  0.5  ,  pe  =  0.3  , 
pc  =  0.2  and  level  of  autocorrelation  q  =  0.2450  (a)  simulation  according  Goodchild's  model  with  1 
random  field,  (b)  sequential  simulation  involving  2  random  fields. 

METHODS  FOR  ERROR  PROPAGATION 

When  source  data  are  input  to  a  GIS  operation  the  uncertainty  which  is  present  in 
the  data  will  be  propagated  through  the  operation  and  will  introduce  uncertainty  in 
the  outcome.  Error  propagation  will  continue  when  the  output  of  one  operation 
becomes  the  input  to  another.  The  network  of  input  and  output  relations  between 
spatial  data  layers  can  be  schematized  in  a  data  flow  diagram  which  is  usually  called 
a  cartographic  model  (Tomlin  1990).  Evaluating  the  accuracy  of  the  output  of  a 
cartographic  model  is  only  feasible  if  the  propagation  of  error  can  be  modelled  for 
each  operation  which  is  part  of  it.  Although  most  GIS  users  may  be  aware  that 
errors  propagate  through  spatial  analysis,  in  practice  they  will  usually  not  pay  any 
attention  to  it.  An  important  reason  for  this  is  the  lack  of  functionality  for  handling 
error  propagation  in  commercial  GIS. 

Lanter  and  Veregin  (1992)  have  proposed  an  automated  system  for  error 
propagation  that  is  based  on  the  use  of  formal  error  propagation  functions  which 
transform  the  PCC  index  of  the  source  layer(s)  through  the  cartographic  model  of 
an  application.  The  problem  with  formal  error  propagation  methods,  however,  is 
that  they  are  only  available  for  a  small  number  of  GIS  operations.  Hence  the  use  of 
tools  which  are  strictly  based  on  formal  error  propagation  methods  will  be  limited  to 


230 


applications  which  only  involve  a  few  simple  GIS  operations.  That  is  why  some 
authors  (Openshaw  1989)  plead  in  favour  of  simulation  modelling  which  puts  no 
restrictions  on  the  complexity  of  the  application,  thereby  arguing  that  the 
computational  load  which  is  the  major  drawback  of  simulation  modelling  will 
become  less  a  problem  with  the  advent  of  increasingly  powerful  means  of 
processing. 

Although  it  may  be  questioned  if  the  use  of  simulation  modelling  will  allow 
enough  runs  to  be  performed  in  real-time  to  obtain  a  statistically  valid  quantitative 
assessment  of  uncertainty  (Heuvelink  1993)  it  will  give  a  clear  indication  of  the 
level  of  uncertainty  in  the  output  which  is  to  be  expected  when  source  data  with 
known  accuracy  are  subjected  to  a  particular  spatial  analysis.  While  formal  error 
propagation  methods  always  produce  a  tabular  report  of  error,  i.e.  a  single  error 
index  or  a  confusion  matrix,  simulation  modelling  has  the  surplus  advantage  of 
providing  interesting  ways  of  visualizing  uncertainty  (see  below).  Hence  it  is  the 
opinion  of  the  present  authors  that  most  advantage  can  be  derived  from  the 
complementary  use  of  both  error  propagation  techniques.  In  the  following  a  short 
description  is  given  of  the  techniques  which  have  been  implemented.  Some 
reference  is  also  made  to  their  practical  use. 


Formal  Error  Propagation  Methods 

Formal  methods  of  error  propagation  can  be  defined  as  mathematical  functions 
describing  how  errors  in  the  source  data  are  modified  by  a  particular  GIS  operation. 
Each  error  propagation  function  is  based  on  a  set  of  assumptions  about  the  nature  of 
the  errors  present  in  the  source  data  (a  particular  model  of  error)  and  is  specific  to 
one  GIS  operation.  Assumptions  about  the  nature  of  error  will  be  based  on  the  type 
of  error  information  that  is  available  for  the  source  data,  i.e.  a  global  error  index, 
class- specific  indices  of  error  or  the  entire  confusion  matrix.  At  present  the 
prototype  tool  only  includes  functions  which  propagate  the  entire  confusion  matrix 
through  the  GIS  operation.  Functions  are  available  for  overlay  (AND,  OR,  XOR) 
and  for  the  reclass  operation  (Veregin  1995).  If  the  entire  confusion  matrix  is  not 
known  it  is  reconstructed  from  available  error  indices  by  making  some  simplifying 
assumptions  about  the  distribution  of  error  among  different  classes.  So  far  no 
assumptions  are  made  on  thematic  dependence,  i.e.  errors  in  different  source  layers 
are  considered  statistically  independent. 

For  many  GIS  operations  little  is  known  about  the  mechanisms  of  error 
propagation  and  formal  error  models  remain  to  be  developed.  Recently  Veregin 
(1994)  reported  on  a  method  to  derive  a  formal  mathematical  expression  describing 
how  error  is  propagated  through  the  buffer  operation  using  simulation  modelling. 
The  method  demonstrates  how  simulation  modelling  can  add  to  a  better 
understanding  of  error  propagation  mechanisms  and  to  the  development  of  formal 
expressions  describing  how  error  is  propagated.  It  provides  a  good  example  of  the 
complementary  use  of  both  error  propagation  techniques  and  of  the  benefits  of 
integrating  both  in  the  same  error  modelling  tool.  Further  research  in  this  direction 
may  lead  to  the  development  of  formal  error  propagation  functions  for  other  more 
complex  GIS  operations. 

Although  the  use  of  formal  methods  for  error  propagation  in  categorical  data  is 
still  limited  to  a  few  GIS  operations  the  technique  is  easy  to  apply.  When 
performing  an  analysis  where  multiple  source  layers  are  subjected  to  a  sequence  of 
GIS  operations  it  may  be  advantageous  to  combine  simulation  modelling  with 


231 


formal  error  propagation  in  order  to  decrease  the  computational  load.  The  combined 
use  of  both  error  propagation  techniques  has  been  a  major  concern  in  designing  the 
tool  which  is  described  in  this  paper. 


Monte  Carlo  Simulation  Modelling 

Simulation  modelling  starts  with  the  definition  of  a  stochastic  error  model  which 
allows  to  produce  a  set  of  perturbed  versions  of  the  original  source  data.  As  already 
explained  the  stochastic  model  will  be  based  on  a  number  of  assumptions  with 
regard  to  the  type,  the  magnitude  and  the  spatial  distribution  of  error.  The  method 
for  stochastic  modelling  which  has  been  presented  in  the  previous  section  and 
which  allows  to  deal  with  autocorrelation  in  the  error  field  has  been  implemented  in 
the  prototype  version  of  the  error  handling  tool.  If  no  information  on  the  spatial 
structure  of  error  is  available  it  is  assumed  that  no  correlation  is  present  in  the  error 
field.  If  the  mean  size  of  inclusions  in  each  class  is  known  sequential  simulation  as 
described  above  can  be  applied. 

Each  perturbed  version  of  the  source  data  is  subjected  to  the  sequence  of  GIS 
operations  involved  in  the  analysis  and  will  produce  a  different  output.  If  the  output 
of  the  analysis  is  a  set  of  numerical  values,  e.g.  the  total  area  allocated  to  different 
types  of  land-use,  the  mean  output  over  M  realizations  gives  an  estimate  of  the  most 
likely  outcome  of  the  analysis  taking  the  uncertainty  in  the  source  data  into  account. 
The  standard  deviation  over  M  realizations  provides  a  measure  of  the  uncertainty  in 
the  output.  If  the  outcome  of  the  analysis  is  an  entire  coverage,  e.g.  a  visibility 
map,  simulation  results  allow  to  generate  a  variety  of  products  describing  output 
uncertainty  (see  below). 

The  implementation  of  the  stochastic  modelling  procedure  is  based  on  scripting. 
When  the  analysis  is  executed  for  the  first  time  the  cartographic  model  is  stored  in  a 
script  file.  Monte  Carlo  analysis  is  automatically  performed  by  running  the  script  a 
fixed  number  of  times.  This  strategy,  of  course,  assumes  that  the  GIS  allows 
scripting  and  that  the  script  language  includes  control  structures  for  iteration. 


VISUALIZATION  OF  UNCERTAINTY 

An  important  issue  in  the  development  of  a  tool  for  error  handling  is  how  to 
visualize  the  quality  of  the  source  data  and  the  reliability  of  the  output  of  spatial 
analysis  in  an  easy  and  straightforward  manner.  To  give  the  user  a  visual 
impression  of  the  error  present  in  the  source  data  it  may  be  interesting  to  visualize 
the  result  of  one  stochastic  simulation  of  error.  When  dealing  with  classified 
satellite  data  class  membership  probabilities  produced  by  statistical  classification 
techniques  can  be  used  to  visualize  uncertainty  which  is  due  to  the  classification 
process.  Based  on  these  probabilities  not  only  the  first,  but  also  the  second,  third, 
...  most  likely  class  can  be  shown.  To  evaluate  the  reliability  of  pixel  assignment  a 
probability  image  can  be  constructed  which  shows  the  highest  class  membership 
probability  for  each  pixel.  Confusion  in  pixel  assignment  can  be  measured  by 
calculating  the  relative  probability  entropy  (Maselli  et  al.  1994).  This  index  indicates 
for  each  pixel  if  probability  tends  to  concentrate  in  one  class  or  is  spread  over 
several  classes. 

When  error  is  propagated  using  the  Monte  Carlo  approach  and  the  outcome  of  the 
analysis  is  an  entire  coverage  the  same  techniques  as  described  for  the  visualization 


232 


of  uncertainty  in  classified  satellite  data  can  be  applied  since  for  each  pixel  a  vector 
indicating  the  frequency  of  occurrence  of  each  class  can  be  constructed  from  the  M 
coverages.  If  error  is  propagated  using  formal  error  propagation  methods 
uncertainty  cannot  be  visualized  since  output  is  in  tabular  form. 


IMPLEMENTATION 

The  above  described  techniques  for  error  modelling  and  error  propagation  have 
been  implemented  in  the  C  programming  language  in  a  UNIX  workstation 
environment.  The  prototype  version  of  the  tool  is  linked  to  GRASS  4.1.  The  choice 
of  GRASS  has  been  based  on  the  fact  that  it  is  public-domain  software,  written  in  C 
and  running  under  all  common  UNIX  operating  systems  and  that  it  provides  a 
powerful  language  for  scripting.  The  most  important  advantage  of  GRASS, 
however,  is  its  open  structure,  meaning  that  developers  can  directly  add  new 
functionality  to  the  system.  The  approach  that  has  been  followed  during  the 
implementation  of  the  error  handling  tool  was  to  consider  the  GIS  as  an  external 
module  which  is  used  to  carry  out  spatial  analysis  and  to  display  maps.  Hence  the 
tool  operates  independently  from  the  GIS  and  should  be  easy-to-link  with  other 
GIS  software. 


CONCLUSION 

This  paper  reviews  existing  methods  for  error  modelling  and  error  propagation  in 
raster-formatted  categorical  data.  The  aim  of  the  work  reported  here  was  to  bundle 
these  methods  in  a  user-friendly  and  independent  tool  for  error  handling  which  can 
easily  be  linked  with  existing  GIS  software.  A  prototype  version  of  the  tool  has 
been  implemented  using  GRASS  4.1  for  raster-based  spatial  analysis  and  display. 

The  tool  combines  the  use  of  formal  error  propagation  methods  with  simulation 
modelling.  Both  techniques  should  not  be  considered  as  two  opposites  but  rather  as 
complementary  techniques.  When  spatial  analysis  involves  the  use  of  GIS 
operations  for  which  no  formal  error  propagation  function  exists  simulation 
modelling  can  be  applied.  At  present  formal  methods  have  been  implemented  for 
overlay  analysis  (AND,  OR,  XOR)  and  for  the  reclass  operation.  They  allow  to 
propagate  the  entire  confusion  matrix  through  the  operation.  Simulation  modelling 
is  based  on  a  method  for  sequential  simulation  which  allows  to  deal  with  spatial 
autocorrelation  in  the  error  field.  In  the  implementation  of  the  method  specific 
measures  were  taken  to  reduce  computational  load. 


ACKNOWLEDGMENTS 

The  work  reported  in  this  paper  is  supported  by  the  Flemish  Institute  for 
Scientific  and  Technological  Research  for  the  Industry  (IWT)  and  by  the  Belgian 
Scientific  Research  Programme  on  Remote  Sensing  by  Satellite  -  phase  three 
(Federal  Office  for  Scientific,  Technical  and  Cultural  Affairs),  contract 
Telsat/III/03/004.  The  scientific  responsability  is  assumed  by  its  authors. 


233 


REFERENCES 


Aspinall,  RJ.  and  D.M.  Pearson,  1994.  A  method  for  describing  data  quality  for 
categorical  maps  in  GIS,  Proc.  of  the  Fifth  European  Conference  and  Exhibition 
on  Geographical  Information  Systems,  EGIS/MARI'94,  Paris,  France,  444-453. 

Congalton,  R.G.  1991.  A  review  of  assessing  the  accuracy  of  classifications  of 
remotely  sensed  data,  Remote  Sensing  of  Environment,  37,  35-46. 

Fisher,  P.F.  1989.  Knowledge-based  approaches  to  determining  and  correcting 
areas  of  unreliability  in  geographic  databases.  Accuracy  of  spatial  databases,  eds: 
Goodchild,  M.  and  S.  Gopal,  45-54. 

Fisher,  P.F.  1991.  Modelling  soil  map-unit  inclusions  by  Monte  Carlo  simulation, 
Int.  J.  of  Geographical  Information  Systems,  5,  2,  193-208. 

Goodchild,  M.F.  and  M.H.  Wang,  1989.  Modelling  errors  for  remotely  sensed 
data  input  to  GIS,  Proc.  on  the  Ninth  International  Symposium  on  Computer- 
Assisted  Cartography,  AUTO-CARTO  9,  Baltimore,  Maryland,  530-537. 

Goodchild,  M.F.,  S.  Guoqing  and  Y.  Shiren,  1992.  Development  and  test  of  an 
error  model  for  categorical  data,  Int.  J.  Geographical  Information  Systems,  6,  2, 
87-104. 

Haining,  R.P.,  D.A.  Griffith  and  R.J.  Bennett,  1983.  Simulating  two-dimensional 

autocorrelated  surfaces.  Geographical  Analysis  15,  247-255. 
Heuvelink,  G.B.M.  1992.  An  iterative  method  for  multidimensional  simulation 

with  nearest  neighbour  models,  in  Dowd,  P.A.  and  J. J.  Royer  (eds.),  2nd 

COD  ATA  Conference  on  Geomathematics  and  Geostatistics,  Nancy,  Sciences  de 

la  Terre,  Series  Informatique  et  Geologique,  31,  51-57. 
Heuvelink,  G.B.M.  1993.  Error  propagation  in  quantitative  spatial  modelling, 

applications  in  geographical  information  systems,  Netherlands  Geographical 

Studies,KNAG,  Faculteit  Ruimtelijke  Wetenschappen  Universiteit  Utrecht,  151  p. 
Lanter,  D.P.  and  H.  Veregin,  1992.  A  research  paradigm  for  propagating  error  in 

layer-based  GIS,  Photogrammetric  Engineering  &  Remote  Sensing,  58,  6,  825- 

833. 

Maselli,  F.,  C.  Conese  and  L.  Petkov,  1994.  Use  of  probability  entropy  for  the 
estimation  and  graphical  representation  of  the  accuracy  of  maximum  likelihood 
classifications,  ISPRS  J.  of  Photogrammetry  &  Remote  Sensing,  49,  2,  13-20. 

Openshaw,  S.  1989.  Learning  to  live  with  errors  in  spatial  databases.  Accuracy  of 
spatial  databases,  eds:  Goodchild,  M.  and  S.  Gopal,  263-276. 

Tomlin,  CD.  1990.  Geographic  Information  Systems  and  Cartographic  Modelling. 
Englewood  Cliffs:  Prentice  Hall. 

Veregin,  H.  1994.  Integration  of  simulation  modelling  and  error  propagation  for  the 
buffer  operation  in  GIS,  Photogrammetric  Engineering  &  Remote  Sensing,  60, 
4,  427-435. 

Veregin,  H.  1995.  Developing  and  testing  of  an  error  propagation  model  for  GIS 
overlay  operations,  Int.  J.  of  Geographical  Information  Systems,  9,  6,  595-619. 


BIOGRAPHICAL  SKETCH 

Both  authors  are  research  assistants  in  the  Department  of  Geography  of  the  Vrije 
Universiteit  Brussel,  Belgium.  Their  main  research  interests  lie  in  spatial  database 
accuracy,  error  modelling  and  error  propagation  in  GIS. 


234 


Discrete  Polygons  or  a  Continuous 
Surface:  Which  is  the  Appropriate 
Way  to  Model  Forests 
Cartographically? 

Kim  Lowell1 


Abstract.— Maps  traditionally  represent  forests  as  a  group  of  discrete 
polygons  —  i.e..  forest  stands.  However,  as  GIS  technology  has 
become  more  common,  there  is  an  increasing  use  of  interpolation 
techniques  to  represent  forest  parameters  as  a  continuous  surface. 
One  must  question  if  both  can  be  correct.  This  paper  presents  the 
results  of  work  conducted  to  assess  the  validity  of  a  polygon  vs. 
continuous  surface  representation  of  forest  volume.  Specifically,  for 
two  different  sample  densities,  spatial  autocorrelation  was  measured 
among  sample  plots.  Following  this  the  improvement  in  volume 
estimation  provided  by  spatial  interpolation  for  individual  plots  was 
assessed  relative  to  the  mean  and  standard  deviation  for  all  plots.  It  is 
concluded  that  in  virtually  no  situation  —  between  forest  types, 
different  stands  of  the  same  forest  type,  and  within  a  single  stand  --  is 
it  appropriate  to  model  forest  volume  as  a  continuous  surface. 
Conversely,  the  evidence  suggests  that  the  forest  is  a  set  of  polygons 
w  ith  inexact  boundaries. 


BACKGROUND  AND  CONTEXT 
Spatial  Interpolation 

Spatial  interpolation  is  a  technique  which  has  existed  since  long  before  digital 
geographic  information  systems  (GIS)  were  created.  It  is  used  to  obtain  estimates 
over  an  entire  surface  for  a  given  variable  which  has  only  been  sampled  at  isolated 
points.  Common  examples  are  estimating  rainfall  and/or  temperature  over  an  entire 
region  for  which  weather  data  were  collected  at  isolated  meteorological  stations.  But 
even  this  seemingly  simple  and  straightforward  example  masks  a  number  of  potential 
problems  with  its  use  in  such  a  situation.  All  are  related  to  two  underlying  implicit 
assumptions  which  must  be  satisfied  if  the  resulting  interpolation  is  to  be  a  valid 
representation  of  the  phenomenon  of  interest. 

First,  the  phenomenon  being  interpolated  must  truly  be  continuous.  Suppose  that 
one  wishes  to  model  rainfall  for  a  region  only  during  a  summer  period  of  hot,  humid 
weather.  In  such  conditions,  it  is  likely  that  the  majority  of  precipitation  will  be 
convective  producing  considerable  amounts  of  rain  in  spot  locations  rather  than  a 
smooth  gradient  of  rainfall  over  a  large  region.  If  this  is  the  case,  then  rainfall  for  the 
summer  period  may  not  be  an  interpolable  phenomenon  due  to  the  seemingly  random 
nature  of  convective  precipitation. 

'Professeur  titulaire.  Centre  de  recherclie  en  geomatique,  Universite  Laval,  Pavilion  Casault,  Ste- 
Foy,  Quebec  G1K  7P4.  Kim.Lowell@for.ulaval.ca.  (418)  656-7998.  FAX:  (418)  656-7411. 


235 


Second,  while  the  initial  necessary  condition  for  interpolability  is  a  continuous 
phenomenon,  the  second  important  factor  is  that  the  phenomenon  being  studied  must 
be  sampled  at  an  appropriate  density  for  its  continuous  nature  to  manifest  itself. 
Suppose  that  in  another  region,  all  precipitation  is  orthographic  —  an  interpolate 
phenomenon  which  is  strongly  related  to  the  topography  of  a  region.  Further 
suppose  that  in  the  region  storm  fronts  arrive  from  the  west  and,  as  they  move  to  the 
east,  encounter  a  series  of  mountain  summits  of  increasing  elevation  evenly  spaced 
10  km  apart.  It  is  reasonable  to  believe  that  the  western  slopes  of  these  mountains 
will  receive  increased  rainfall  as  one  moves  towards  a  summit.  Moreover  the  amount 
of  rainfall  will  fall  off  sharply  as  a  summit  is  crested  given  that  a  storm  rises  in 
contact  with  a  mountain  slope,  cools,  releases  all  possible  precipitation,  and  moves 
towards  the  next  higher  mountain  where  the  process  repeats  itself.  In  this 
hypothetical  case,  if  weather  stations  are  spaced,  say,  40  km  apart,  the  resulting  data 
for  rainfall  may  not  lend  themselves  to  being  interpolated  even  though  the 
phenomenon  being  studied  is  known  to  be  interpolate.  A  better  approach  might  be 
a  regression  model  in  which  the  amount  of  rainfall  is  estimated  as  a  function  of 
elevation  and  west-  or  east-facing  slope. 

Thus  the  ability  to  conduct  reliable  spatial  interpolation  depends  both  on  the  nature 
of  the  phenomenon  being  studied  and  also  the  way  in  which  it  has  been  sampled. 

Forest  Parameters 

Traditionally  forest  maps  are  composed  of  non-overlapping  polygons  having  non- 
overlapping  appellations  within  the  classification  system  used.  Yet  recently  — 
particularly  with  the  advent  of  GIS  and  associated  technologies  ~  interpolation  has 
begun  to  be  used  to  produce  continuous  surfaces  from  forest  sample  points  for 
various  forest  parameters. 

And,  indeed,  there  may  be  reason  to  believe  that  at  some  scale  forest  parameters  are 
interpolate  phenomenon.  Given  a  single  forest  stand,  for  example,  it  may  be  that  "a 
lot"  of  its  volume  is  concentrated  in  a  relatively  small  area  with  gradations  towards 
adjoining  forest  stands.  Some  researchers  (Joy  and  Klinkenberg  1994)  have  noted 
the  tendency  for  forest  polygon  cores  to  be  the  most  precise  —  i.e.,  repeatable  by 
several  interpreters  —  part  of  a  polygon  suggesting  that  forest  parameters  may  show  a 
gradient  from  polygon  centre  to  polygon  edges.  Such  information  has  led  to  at  least 
one  representation  of  spatial  certainty  for  various  forest  types  based  on  the  idea  that 
the  polygon  centre  is  the  part  of  the  polygon  which  is  most  likely  to  truly  be  the 
forest  type  labelled  —  an  interpolation  of  certainty  (Lowell  1994).  And  between 
forest  stands  of  the  same  type  or  even  different  forest  types  it  may  be  possible  that 
forest  parameters  are  interpolate.  A  pure  spruce  stand  in  northern  Quebec  may 
contain  less  volume  than  the  same  stand  type  in  southern  Quebec.  Similarly,  a  high 
volume  stand  in  one  location  may  suggest  that  surrounding  stands  will  also  have 
high  volumes  due  to  favourable  conditions  related  to  underlying  factors  such  as  soil 
type  and  water  availability. 

The  converse  may  also  be  true,  however.  Consider  that  the  spatial  structure  of 
within-stand  variance  is  not  known  and  would  seem  to  depend  strongly  on  the  forest 
sampling  unit  employed  —  a  smaller  sample  unit  giving  a  larger  variance  than  a  larger 
unit.  Moreover,  in  certain  stands  the  spatial  variability  of  tree  locations  and  the 
characteristics  of  the  sample  units  employed  may  render  the  spatial  structure  of  the 
variance  seemingly  random  —  i.e.,  non-interpolable.  Similarly,  if  a  forest  has  been 
subjected  to  periodic  disturbances  such  as  fires  or  severe  wind,  the  characteristics  of 
one  stand  may  provide  absolutely  no  indication  of  the  characteristics  of  nearby 


236 


stands.  Moreover,  even  different  forest  stands  within  the  same  forest  type  may 
show  no  gradation  of  productivity.  Over  a  long  range  —  i.e.,  northern  to  southern 
Quebec  —  it  may  not  be  possible  to  have  the  same  forest  type  appear  over  such  a 
range  due  to  biological  carrying  capacities  of  various  sites.  And  even  over  a  smaller 
region  the  productivity  of  the  stands  of  a  given  forest  type  may  be  more  related  to 
local  ecophysiographic  conditions  than  to  a  mere  gradation  of  productivity. 

Objectives 

The  purpose  of  this  paper  is  to  examine  the  question  of  whether  or  not  forest  volume 
is  an  interpolable  phenomenon  in  the  boreal  forest  of  Quebec.  This  question  will  be 
addressed  at  a  number  of  levels  as  it  concerns  volume  among  forest  stands  of 
different  forest  types,  different  forest  stands  in  the  same  forest  type,  and  within 
individual  forest  stands.  In  addition  to  assessing  whether  or  not  volume  is 
interpolable  at  any  of  these  scales,  the  implications  of  conducting  interpolation  when 
it  is  not  warranted  for  these  conditions  will  also  be  evaluated  and  presented. 

STUDY  AREA,  DATA,  AND  METHODS 

Two  real  world  forestry  data  sets  were  employed  in  this  study.  Both  contain  the 
per  hectare  wood  volumes  estimated  for  a  collection  of  0.04  ha  sample  plots  located 
in  Montmorency  Forest  —  the  research  forest  of  Laval  University  located  80  km 
north  of  Quebec  City.  The  dominant  vegetation  on  Montmorency  Forest  is  the  boreal 
forest  dominated  by  balsam  fir  (Abies  balsamea  (L.)  Mill.)  and  to  a  lesser  extent 
spruce  (Picea  spp.)  and  white  birch  (Betula  papyrifera  Marsh.)  (Belanger  et  al. 
1988).  The  first  data  set  is  composed  of  92  sample  plots  distributed  in  a  semi- 
random  stratified  sample  throughout  the  1500  ha  study  area.  These  plots  had  been 
established  as  part  of  normal  management  activities  to  assess  the  harvestable  volume 
on  the  forest.  They  were  therefore  located  in  mature  stands  or  those  approaching  a 
harvestable  volume.  The  density  of  these  plots  (1  per  16  ha  or  1  plot  every  400  m  on 
average)  is  approximately  20  times  as  dense  as  forest  inventories  normally  taken  in 
Quebec  (Beaulieu  and  Lowell  1994).  The  second  data  set  contains  81  plots  located 
within  a  16  ha  square  in  Montmorency  Forest.  These  plots  were  placed  on  a  9-by-9 
grid  with  plots  being  located  50  m  apart.  The  16  ha  grid  was  located  intentionally 
such  that  a  number  of  forest  types  covering  a  range  of  forest  conditions  were 
included  in  the  sample:  clearcuts,  partial  and  total  windthrow,  and  mature  and 
developing  forests. 

Two  types  of  analyses  were  conducted  on  each  data  set  in  its  entirety  and  a  series 
of  subsets.  These  subsets  will  be  described  subsequently. 

First,  to  assess  the  interpolability  of  volume,  the  spatial  autocorrelation  was 
measured  among  the  plots  of  a  given  data  (sub)set  (see  Table  1).  This  was  assessed 
by  using  both  Moran's  /  and  Geary's  c  (Griffith  1985)  and  the  assumption  that 
sample  plots  were  adjacent  if,  in  a  Voronoi  diagram  constructed  around  the  plots, 
their  Thiessen  polygon's  touched.  If  a  phenomenon  is  interpolable,  it  is  expected 
that  high  values  will  have  a  tendency  to  cluster  in  the  same  area  as  will  low  values. 
This  is  the  same  as  saying  that  positive  spatial  autocorrelation  will  be  present.  T 
values  for  /  and  c  are  presented  to  indicate  significant  differences  from  zero.  If  forest 
volume  is  interpolable,  it  is  expected  that  these  values  will  be  positive  and 
significant. 

Second,  a  given  data  (sub)set  was  divided  into  a  calibration  and  validation  subset. 
The  validation  subset  contained  a  random  sample  of  interior  plots  that  was 


237 


approximately  one-third  of  all  the  interior  plots.  The  calibration  data  set  contained 
the  remaining  plots.  Spatial  interpolation  using  Voronoi  diagram  area  stealing  (Gold 
1989)  was  conducted  on  the  calibration  data  set  and  the  interpolated  volumes  for  each 
of  the  validation  data  plot  locations  extracted.  The  mean  difference  of  the  true  and 
interpolated  values  was  calculated  and  tested  for  a  significant  difference  from  zero. 
The  two  values  were  also  regressed  against  each  other;  perfect  agreement  will 
provide  an  R2  of  1 .0,  a  slope  of  1 .0  and  an  intercept  of  0.0,  a  root-mean-square- 
error  (RMSE)  of  0.0. 


Table  1.  Data  sets  and  subsets  treated,  motivation   for  subset,  and  number  of 

points  in  the  validation  and  calibration  subsets  of  each.  

 Data  (Sub)set  Motivation  Points:  Total  Cal  Val 

J 500  ha  (Between  stand  interpolability) 

All  plots                      Global  interpolability  92  77  15 

Softwood  Plots              Interpolability  within  species  81  69  12 

Fir  plots  >60%  density     Interpolability  of  fully  stocked  stands  61  52  9 

16  ha  (Within  stand  interpolability) 

All  plots                      Global  interpolability  81  67  14 

>  250  stems  per  ha         Across  species  types  w/o  clearcuts  65  54  11 

and  blowdowns 

Softwood  and  >              Within  a  single  (dominant)  species  type  55  46  9 
250  stems  per  ha 
Benchmarks 

Topographic                  Known  interpolate  phenomenon  1037  998  39 
Random                       Non-interpolable  phenomenon  100  81  19 
 Synthetic  data  Explanation  of  results  36  29  7 

The  subsets  of  the  two  real  world  forestry  data  sets  examined  were  designed  to  test 
the  interpolability  of  forest  volume  at  a  variety  of  levels  under  the  conditions  studied 
(Table  1).  For  the  92  plots  over  1500  ha  the  subsets  were  a)all  data  together,  b)only 
plots  labelled  as  softwood  on  the  forest  map  of  the  area,  and  c)only  plots  labelled 
predominantly  fir  with  a  density  greater  than  60%.  For  the  81  plots  on  a  16  ha  grid 
the  subsets  were  a)all  data  together,  b)only  plots  having  250  stems  per  ha  and 
higher,  and  c)only  softwood  plots  having  250  stems  per  ha  or  higher. 

For  a  point  of  comparison,  three  additional  data  sets  were  also  treated  and  analysed 
in  exactly  the  same  manner  as  the  other  data  (sub)sets.  First,  a  digital  terrain  model 
for  a  1  km  (100  ha)  square  developed  for  another  study  was  employed  (Lowell 
1994).  Second,  100  locations  were  randomly  generated  over  a  500  m  square  area 
(25  ha)  and  a  value  between  1  and  100  (with  replacement)  randomly  assigned  to  each 
of  the  locations.  These  two  data  sets  provide  benchmarks  for  what  one  should  find 
for  a  phenomenon  known  to  be  interpolate  (elevation)  and  one  which  is  not  at  all 
expected  to  be  interpolate  (random  locations  and  attribute  assignments).  The  third 
synthetic  data  set  was  a  grid  of  36  points  spaced  50  m  apart  (Figure  1).  The  use  of 
this  data  set  will  become  clear  after  the  results  of  the  other  data  sets  have  been 
presented  and  discussed.  For  the  moment,  suffice  it  to  say  that  this  data  set  is 
designed  to  simulate  three  forest  types  which  were  sampled  systematically  and 
densely.  The  value  assigned  at  each  location  was  random  within  the  range  assigned 
a  priori  to  each  type:  A)  10  to  40,  B)  30  to  70,  and  Q60-90.  The  overlap  between 
classes  is  intentional  as  it  reflects  real  forest  conditions  wherein  it  is  possible  that  the 
volume  at  any  given  location  for  a  particular  type  is  less  than  that  found  in  a  type 


238 


known  to  contain  less  volume.  The  mean,  of  course,  of  the  two  types  will  reflect 
their  relative  densities. 


RESULTS  AND  DISCUSSION 

The  results  for  the  two  benchmark  data  sets  provide  a  useful  manner  in  which  to 
assess  the  results  for  the  real  world  forest  data  sets  (Table  2).  In  general,  both  the 
topographic  and  the  random  data  sets  perform  as  expected. 

The  topographic  data  are  strongly  positively  spatially  autocorrelated  suggesting  that 
interpolation  for  this  phenomenon  at  the  scale  sampled  is  warranted.  This  is  further 
reinforced  by  noting  the  strong  relationship  that  exists  between  the  true  and 
interpolated  data  for  the  validation  data  points  (R2=0.97),  and  the  mean  difference 
between  the  two  sets  of  values  is  fairly  small.  Moreover,  the  slope  and  intercept  of 
the  regression  line  are  close  to  the  expected  values  of  1.0  and  0.0,  respectively. 
Another  measure  of  the  usefulness  of  interpolation  can  be  seen  in  the  value  of  the 
RMSE  (2.5)  compared  to  that  of  the  standard  deviation  (14.1)  for  all  data.  In  the 
absence  of  the  interpolated  values,  the  best  estimate  for  a  given  point's  elevation  will 
be  60.1  (the  overall  mean)  with  the  error  varying  ±  28  with  95%  confidence  (a  95% 
confidence  limit  using  the  standard  deviation  under  an  assumption  of  a  normal 
distribution).  The  interpolation  will  provide  an  estimate  at  a  given  location  that  varies 
only  +  5  with  95%  confidence. 

By  contrast,  the  random  data  set  indicates  no  spatial  autocorrelation  structure  (non- 
significant /  and  c)  suggesting  that  spatial  interpolation  is  not  warranted.  And  indeed 
the  R2  indicates  that  no  correlation  exists  among  the  interpolated  and  true  values  for 
the  validation  data  points.  Furthermore  the  slope  of  the  line  is  significantly  different 
from  the  expected  value  of  1 .0  though  not  significantly  different  from  a  slope  of  zero 
(which  indicates  no  relation).  In  the  absence  of  a  useful  interpolation,  the  value 
estimated  at  a  given  point  will  be  the  mean  (52.6)  +  56. 1  with  95%  confidence  while 
the  interpolation  gives  estimates  within  61.9  units  with  95%  confidence.  Clearly,  the 
random  values  should  not  have  been  interpolated  as  indicated  by  the  spatial 
autocorrelation  statistics  and  the  subsequent  regression  analysis.     In  fact,  the 


239 


consequences  of  having  conducted  an  interpolation  when  it  was  not  warranted  has 
been  in  increase  in  the  error  on  the  resulting  surface.  In  a  practical  sense,  this  means 
that  the  resulting  surface  has  virtually  no  accuracy  for  a  given  location. 

Table  2.  Results  of  analysis  for  spatial  autocorrelation  and  interpolation. 

Mean1  Spat.  Auto.2   Regression3   All  Data4 

Data  set          Diff.       /         c         R2    Slope    Intercept    RMSE  Mean  Std  Dev. 

1 500  ha  (Between  stand  Interpol  ability) 

All  plots        -18.6    0.98     0.62      0.16    0.95       -12.5           65.5  118.6  62.8 

Softwood  Plots  36.5    0.97     0.93      0.00   -0.38*     139.1           70.1  116.8  63.8 

Fir  plots  >      -30.8     0.53     0.41      0.00    0.08**    90.0           62.1  121.3  62.2 
60%  density 

16  ha  (Within  stand  interpolability) 

All  plots           2.5     4.02**  3.56**    0.00    0.14        66.9           41.7  73.7  43.2 

>  250  stems/ha  2.1     3.25**  1.88      0.00    0.38        54.0*         22.3  89.5  31.6 

Softwood  and  -11.9     1.77     0.81       0.00  -0.24        97.3**        28.2  92.6  31.7 
>  250  stems/ha 

Benchmarks 

Topographic    -0.1    52.26**  47.36**  0.97  1.04          2.4            2.5  60.1  14.1 

Random         -5.9   -1.01      -0.38      0.00  0.18*       39.6          31.6  52.6  28.6 

Synthetic  Data -1.3     3.16**     2.02*    0.07  1.22        -12.7          28.2  48.4  19.9 

'Validation  data  only.  "*"/"**"  indicates  significant  difference  from  zero  at  95%/99%. 

2t  values  for  Moran's  /  and  Geary's  c.  for  all  data.  "*"/"**"  indicates  significant  difference  from  zero 
at  95%/99%. 

3  Validation  plots  only.  "*"/"**"  indicates  significant  difference  from  expected  values  at  95%/99%. 

4  All  data  for  purposes  of  comparison. 

It  is  also  apparent  that  none  of  the  situations  explored  using  the  92  sample  plots  are 
interpolate.  Perhaps  the  quickest  way  to  see  this  is  to  note  that  none  of  the 
interpolations  provided  RMSEs  that  were  noticeably  different  from  the  standard 
deviations  of  the  data.  However,  this  is  also  abundantly  clear  in  the  low  R2s,  the 
coefficients  of  the  regression  lines,  and  the  lack  of  significance  of  either  spatial 
autocorrelation  coefficient.  The  importance  of  these  findings  is  considerable. 

This  demonstrates  that,  in  the  conditions  studied,  between-stand  interpolation  is 
not  warranted.  This  is  perhaps  not  surprising  in  the  case  of  global  interpolability  (All 
plots)  and  interpolability  within  species  (Softwood  plots)  considering  that 
Montmorency  Forest  is  a  managed  forest  on  which  normal  harvesting  activities  have 
been  ongoing.  Thus  the  volume  of  the  stands  represented  in  these  two  cases  may 
reflect  more  on  past  management  practices  than  the  biological  productivity  of  these 
areas  —  even  though  plots  were  located  in  relatively  mature  forests.  Note  that  these 
conclusions  give  little  idea  about  what  would  be  found  in  forests  which  have  not 
been  subjected  to  wide-spread  natural  or  human  disturbances.  However,  it  is  argued 
here  that  such  forests  are  not  very  common  meaning  that  these  results  should  be 
widely  applicable. 

The  findings  for  the  92  plots  also  indicate  that  one  cannot  interpolate  even  for  fully 
stocked  stands  of  similar  species  combination  which  have  not  been  subjected  to 
recent  disturbance.  These  stands  (Fir  plots  >  60%  density)  should  have  had  time  to 
manifest  their  biological  potential  and  possibly  show  a  gradient  of  productivity 
across  the  roughly  7  km  by  3  km  study  area.  It  might  be  suggested  that  the  reason 


240 


that  this  was  not  the  case  was  that  the  productivity  of  each  stand  is  related  to  the 
underlying  ecophysiography  of  a  stand  which  was  not  considered  here.  However, 
in  another  study  which  used  the  same  92  sample  plots  (Coulombe  and  Lowell  1995) 
the  relation  of  the  productivity  of  each  site  to  its  underlying  ecophysiography  as 
estimated  by  map-based  data  (soils,  slope,  etc.)  was  examined;  no  significant 
relationships  were  found.  Thus  it  would  appear  that  over  a  large  area,  even  with  a 
sampling  density  much  higher  than  is  usually  employed,  interpolation  of  forest 
volume  is  not  appropriate  even  for  stands  which  have  remained  undisturbed  for  a 
long  period. 

At  first  glance,  the  results  for  the  various  subsets  of  the  81  sample  plots  might 
indicate  that  within-stand  interpolation  is  possible  —  both  spatial  autocorrelation 
coefficients  are  positive  and  relatively  large  for  All  plots  and  plots  having  >250  stems 
per  ha.  However,  the  associated  R  s  are  zero  (0.0)  and  the  RMSEs  are  not  largely 
different  from  the  standard  deviations.  A  visual  examination  of  the  data  makes  it 
very  clear  that  the  reason  the  spatial  autocorrelation  statistics  are  strong  in  the  case  of 
All  Plots  is  the  presence  of  four  adjacent  plots  being  located  in  a  clearcut,  and  another 
five  plots  being  located  in  a  blowdown  area.  Simply  stated,  low  volumes  cluster  in 
these  areas  —  i.e.,  are  positively  spatially  autocorrelated.  When  these  are  removed, 
spatial  autocorrelation  decreases,  though  it  remains  positive  and  strong.  While  this 
suggests  that  interpolation  would  be  appropriate,  the  regression  statistics  suggest 
otherwise. 

This  can  be  explained  by  considering  the  synthetic  surface  in  Figure  1  and  its 
associated  statistics  in  Table  1 .  This  surface  was  constructed  so  that  large  values 
would  have  a  tendency  to  cluster  (which  is  the  case  as  measured  by  the  spatial 
autocorrelation  coefficients  for  this  surface).  But  if  one  removes  a  given  observation 
and  attempts  to  estimate  its  value  from  surrounding  observations,  the  precision  of 
estimates  is  fairly  low.  That  is,  like-values  show  tendencies  to  group,  but  within 
each  group  there  is  no  consistency  of  estimates.  This  suggests  that  there  are 
identifiable  areas  —  i.e.,  polygons  —  of  high  and  low  values,  but  within  those  areas, 
the  variance  is  spatially  random. 

That  this  is  the  case  in  real  forests  is  further  reinforced  by  considering  the 
Softwood  and  >  250  stems  per  ha  plots.  In  this  data  subset,  spatial  autocorrelation  is 
weakened  to  the  point  where  it  is  no  longer  statistically  significant,  and  the 
regression  statistics  show  that  interpolation  has  not  provided  better  estimates  than 
would  the  mean  and  the  standard  deviation  alone.  Given  that  the  case  of  Softwood 
and  >  250  stems  per  ha  can  be  considered  a  single  type,  if  one  cannot  observe  a 
spatial  tendency  of  volumes  within  a  stand  then  one  should  not  conduct  within-stand 
interpolation. 

To  understand  all  of  the  findings  of  this  study  —  i.e.,  a  complete  lack  of 
interpolability  —  one  must  consider  the  nature  of  forest  sampling.  With  conventional 
forest  inventory  techniques,  the  volume  found  at  any  given  location  is  somewhat  a 
function  of  the  micro-location  of  the  plot.  That  is,  each  forest  type  has  an  associated 
range  of  volume  that  one  will  find  at  any  given  location.  When  one  samples  a  given 
location,  one  cannot  know  if  one  is  at  the  upper,  lower,  or  middle  portion  of  that 
range.  Thus  a  large  part  of  the  reason  that  between-  and  within-stand  interpolability 
is  inappropriate  is  simply  the  natural  random  spatial  variability  of  a  forest  type  and 
the  way  in  which  it  is  sampled.  One  could  minimise  such  effects  by  using  a  larger 
sample  unit  so  that  one  is  consistently  sampling  the  mean  of  a  type  (or  the  range  for 
the  type  would  be  reduced).  But  doing  so  risks  to  require  an  excessive  amount  of 
field  time  and  also  to  exceed  the  boundaries  of  the  forest  stand  being  sampled. 


241 


CONCLUSIONS 


It  appears  that  it  is  not  appropriate  to  model  the  forest  as  a  continuous  surface. 
Instead,  the  results  from  this  study  suggest  that  the  forest  is  a  set  of  polygons.  Note, 
however,  that  this  does  not  mean  that  one  has  polygons  with  crisp  boundaries  as  is 
usually  represented  on  forest  maps.  While  the  synthetic  surface  employed  herein 
(Figure  1)  was  based  on  definite,  widthless  boundaries,  a  number  of  different 
interpreters  presented  with  only  the  36  points  in  Figure  1  would  probably  not 
develop  the  same  map.  And  even  if  such  interpreters  were  provided  with  the 
knowledge  that  there  were  only  "truly"  three  types  present,  it  is  doubtful  that  the 
lines  would  be  placed  in  the  same  locations  by  different  interpreters.  Thus  the  best 
interpretation  of  how  to  represent  a  forest  might  be  as  a  set  of  polygons  whose 
boundaries  show  a  range  of  possible  locations. 

ACKNOWLEDGEMENTS 

The  financial  support  for  this  work  provided  by  the  Association  of  Quebec  Forest 
Industries  and  the  Natural  Sciences  and  Engineering  Research  Council  of  Canada  is 
gratefully  acknowledged. 

BIBLIOGRAPHY 

Beaulieu,  P.,  Lowell,  K.,    1994.  Spatial  autocorrelation  among  forest  stands 

identified  from  the  interpretation  of  aerial  photographs.  Landscape  and  Urban 

Planning  29:161-169. 
Belanger,  L.,  Bertrand,  L.,  Bouliane,  P.,  Lussier,  L.J.,  1988.  Plan  d'amenagement 

de  la  foret  Montmorency.     Universite  Laval,  Faculte  de  foresterie  et  de 

geomatique,  215  pp. 

Coulombe,  S.,  Lowell,  K.,  1995.  Ground-truth  verification  of  relations  between 

forest  basal  area  and  certain  ecophysiographic  factors  using  a  geographic 

information  system.  Landscape  and  Urban  Planning  32: 127-136. 
Gold,  CM.,  1989.  Surface  interpolation,  spatial  adjacency  and  G.I.S.  Chapter  3  in 

Three  dimensional  applications  in  geographic  information  systems  (J.  Raper, 

Editor),  Taylor  and  Francis,  London,  pp.  21-35. 
Griffith,  D.A.,  1985.  Spatial  autocorrelation:  a  primer.  Monograph  of  the  American 

Association  of  Geographers,  Washington,  82  pp. 
Joy,  M.,  Klinkenberg,  M.,  1994.    Proceedings:  GIS  "94,  February,  Vancouver, 

British  Columbia,  pp.  359-365. 
Lowell,  K.E.,  1994.  Probabilistic  temporal  GIS  modelling  involving  more  than  two 

map  classes.  International  Journal  of  Geographical  Information  Systems  8:73-93. 
Lowell,  K.E.,  1994.  fuzzy  cartographic  representation  for  forestry  based  on  Voronoi 

diagram  area  stealing.  Canadian  Journal  of  Forest  Research  24: 1 970- 1 980. 
Odland,  J.,  1988.  Spatial  Autocorrelation.  Scientific  Geography  Series,  Volume  9, 

Sage  Publications,  Beverly  Hills,  California.  87  pp. 

BIOGRAPHICAL  SKETCH 

Kim  Lowell  is  a  Full  Professor  of  Forestry  Geomatics  at  Laval  University  in 
Quebec  City,  Canada  within  the  Industrial  Research  Chair  in  Geomatics  Applied  to 
Forestry.  He  received  a  Ph.D.  in  Forest  Biometrics  from  Canterbury  University, 
New  Zealand  in  1985. 


242 


Choosing  between  Abrupt  and 
Gradual  Spatial  Variation? 

Gerard  B.M.  Heuvelink1  and  Johan  A.  Huisman2 


Abstract. -Two  basic  models  of  spatial  variation  are  widely  used  in 
present-day  soil  survey  practice.  The  discrete  model  of  spatial  variation 
(DMSV)  forms  the  basis  of  the  traditional  soil  map,  in  which 
homogeneous  soil  mapping  units  are  separated  by  abrupt  boundaries.  The 
continuous  model  of  spatial  variation  (CMSV)  originates  from 
geostatistics,  where  kriging  is  used  to  map  gradual  changes  in  soil 
properties.  Neither  of  the  two  models  is  capable  of  handling  situations  in 
which  abrupt  and  gradual  spatial  variation  are  both  present  in  the  same 
area.  Therefore,  recently  a  straightforward  combination  of  the  two 
models  was  introduced,  known  as  the  mixed  model  of  spatial  variation 
(MMSV).  The  MMSV  contains  the  two  basic  models,  suggesting  that  it 
may  perform  well  on  the  whole  range  of  spatial  variation.  In  this  paper 
we  investigate  the  anticipated  flexibility  of  the  MMSV  using  a  simulation 
study.  As  expected,  the  MMSV  is  superior  to  the  DMSV  and  the  CMSV 
when  both  types  of  spatial  variation  are  present.  But  the  MMSV  is  also 
as  suitable  as  the  DMSV  in  case  of  a  'discrete'  reality,  and  as  suitable  as 
the  CMSV  in  case  of  a  'continuous'  reality.  From  this  we  conclude  that 
the  MMSV  should  be  recommended  for  situations  where  an  a  priori 
choice  between  abrupt  and  gradual  spatial  variation  cannot  easily  be 
made. 


INTRODUCTION 

Traditional  soil  survey  is  based  upon  the  presumption  that  soil  behaves 
uniformly  within  soil  mapping  units  and  changes  fairly  abrupt  at  the  boundaries 
between  them  (Voltz  and  Webster  1990,  Webster  and  Oliver  1990,  Burrough 
1993).  However,  the  practical  validity  of  this  conventional  representation  of  soil 
variability  has  repeatedly  been  questioned  (e.g.  Webster  and  Cuanalo  1975, 
Nortcliff  1978,  Campbell  et  al.  1989,  Nettleton  et  al.  1991).  Major  drawbacks 
of  the  conventional  model  of  soil  spatial  variation  are  that  it  cannot  represent 
gradual  boundaries  and  that  it  ignores  spatial  autocorrelation  within  mapping 
units. 

As  an  alternative  to  the  conventional  model,  some  fifteen  years  ago  the  use 
of  geostatistical  techniques  for  modelling  spatial  variation  was  introduced  to  soil 
science.  Burgess  and  Webster  (1980)  were  among  the  first  to  apply  kriging,  an 


'Geostatistician,  University  of  Amsterdam,  The  Netherlands 
2 Student,  University  of  Amsterdam,  The  Netherlands 


243 


exponent  of  this  theory,  to  soil  survey.  Since  then  the  geostatistical  approach  to 
modelling  soil  spatial  variation  has  flourished  and  kriging  is  now  routinely  being 
adopted  for  the  mapping  of  soil  properties  (Webster  1994). 

Recently,  however,  it  is  more  and  more  being  realized  that  the  outright 
abandoning  of  the  conventional  approach  to  soil  spatial  variation  is  perhaps  too 
drastic.  Kriging  definitely  has  disadvantages  as  well,  such  as  its  inadequacy  to 
deal  with  sharp  boundaries.  In  order  to  bridge  the  gap  between  the  conventional 
and  geostatistical  representation  of  soil  spatial  variation,  several  models  have 
been  developed  that  can  handle  discrete  (abrupt)  as  well  as  continuous  (gradual) 
spatial  variation  in  the  same  area  (Stein  et  al.  1988,  Voltz  and  Webster  1990, 
Heuvelink  and  Bierkens  1992,  Rogowski  and  Wolf  1994,  Goovaerts  and  Journel 
1995,  Heuvelink  1996). 

In  this  paper  we  examine  one  such  a  model,  known  as  the  mixed  model  of 
spatial  variation  (MMSV).  In  Heuvelink  (1996)  it  was  anticipated  that  the 
MMSV  should  work  well  on  the  whole  range  of  spatial  variation,  from  purely 
discrete  to  purely  continuous.  We  analyze  the  anticipated  flexibility  of  the 
MMSV  using  nine  simulated  'realities'.  But  before  describing  the  exact 
procedure  of  the  simulation  exercise,  we  will  first  briefly  review  the  three 
models  of  spatial  variation  used  in  this  study. 

THREE  MODELS  OF  SPATIAL  VARIATION 

The  Discrete  Model  of  Spatial  Variation  (DMSV)  first  divides  the 
geographical  domain  D  into  K  separate  units  Dk.  It  then  makes  the  following 
assumptions  on  the  behaviour  of  a  spatially  distributed  attribute  Z(-): 

1  Z(x) =/ik + e(x)  for  all  x  E  Dk 

2  e(-)  has  zero  mean  and  is  spatially  uncorrelated 

3  Var(e(x))  =  C0  for  all  x  E  D 

Thus  the  DMSV  assumes  that  Z(x)  is  the  sum  of  a  unit-dependent  mean  fik  and  a 
residual  noise  e(x).  The  DMSV  will  usually  be  adopted  when  the  units  Dk  are 
available  in  the  form  of  a  polygon  map,  such  as  a  soil  map,  a  landuse  map  or  a 
geological  map,  and  when  the  within-unit  variability  is  expected  to  be  small  in 
comparison  with  the  between-unit  variability.  In  other  words,  the  DMSV 
represents  the  conventional  model  of  soil  spatial  variation  and  is  appropriate 
when  major  jumps  in  the  attribute  Z(  )  take  place  at  the  boundaries  of  the  map- 
ping units. 

In  its  simplest  form,  the  Continuous  Model  of  Spatial  Variation  (CMSV) 
makes  the  following  assumptions: 

1  E[Z(x)]=^forallxED 

2  Cov(Z(x),Z(x+h))=Cz(|h|)  for  all  x,x+hED 


244 


Thus  the  CMSV  assumes  that  Z(  )  is  second-order  stationary,  meaning  that  it 
has  a  constant  mean  and  that  its  spatial  autocovariance  is  a  function  only  of  the 
distance  between  the  locations.  The  CMSV  embodies  the  geostatistical  model  of 
spatial  variation.  In  geostatistics  it  is  customary  to  use  the  variogram  yz(-)  to 
characterize  the  spatial  autocorrelation  of  Z(  ).  It  is  related  to  the  autocovariance 
by  the  identity  yz(  |  h  | )  =  Cz(0)-Cz(  |  h  | ) . 

The  assumptions  underlying  the  Mixed  Model  of  Spatial  Variation  (MMSV) 
are  a  combination  of  those  underlying  the  DMSV  and  CMSV: 

1  Z(x)=/xk  +  e(x)  for  all  x(EDk 

2  e(-)  has  zero  mean 

3  Cov(e(x),e(x  +  h))  =  Cf(|n|)  for  all  x,x+hED 

Second-order  stationarity  is  thus  imposed  on  e(-)  instead  of  on  Z(-).  The  MMSV 
is  more  general  than  the  DMSV  and  CMSV,  and  in  fact  it  contains  both  models. 
However,  the  DMSV  and  CMSV  are  included  here  as  separate  models  because 
they  are  very  often  used  in  practice. 

APPLICATION  TO  NINE  SIMULATED  REALITIES 

In  order  to  study  the  suitability  of  the  three  MSV's  for  mapping  under 
various  circumstances,  nine  different  'realities'  were  created.  This  was  done  by 
adding  maps  generated  using  unconditional  Gaussian  simulation  (Deutsch  and 
Journel  1992)  to  an  artificially  constructed  'soil  map'.  The  nine  realities  are 
given  in  figure  1.  The  A-maps  in  figure  1  (top  row)  are  strongly  dominated  by 
the  discrete  soil  map,  the  B-maps  (middle  row)  to  a  much  lesser  extent,  and  the 
C-maps  (bottom  row)  bear  no  influence  from  the  discrete  soil  map.  The  degree 
of  spatial  autocorrelation  in  the  added  residual  decreases  from  the  1-maps  (left 
column)  to  the  3-maps  (right  column). 

Mapping  the  soil  property  from  200  observations 

From  each  of  the  nine  simulated  maps  data  sets  were  created  by  collecting 
observations  at  200  randomly  selected  locations.  From  these  observations  the 
soil  property  was  mapped  using  the  three  MSV's.  This  resulted  in  27  maps  of 
predictions  and  27  maps  of  prediction  error  standard  deviations. 

Mapping  using  the  DMSV  is  done  simply  by  calculating  the  mean  of  the 
observations  for  all  mapping  units  separately,  and  using  the  unit  mean  as  a 
prediction  for  all  points  lying  in  the  same  unit.  Mapping  with  the  CMSV  is  done 
by  ordinary  kriging,  in  which  a  variogram  is  used  computed  from  the  200 
observations.  Mapping  is  somewhat  more  complicated  in  case  of  the  MMSV. 
First  a  variogram  is  computed  from  the  200  residuals  obtained  by  subtracting  the 
unit  means  from  the  observations  (Kitanidis  1994).  Next  the  attribute  is  mapped 
using  universal  kriging  (Cressie  1991). 


245 


Al  A2  A3 


Figure  l.-Nine  simulated  realities.  Letter  indicates  influence  from  the  artificial  soil 
map:  A  =  large,  B= moderate,  C  =  none.  Number  indicates  degree  of  spatial 
autocorrelation  in  the  added  residual:  1  =  large,  2= moderate,  3= none. 


In  figure  2  the  results  of  the  mapping  are  given  for  a  selection  of  three  out 
of  the  nine  simulated  realities.  Note  that  these  are  the  prediction  maps,  and  that 
the  corresponding  maps  of  prediction  error  standard  deviations  are  not  given 
here.  The  DMSV  maps  necessarily  follow  the  delineations  of  the  soil  map, 
which  is  quite  all  right  for  map  A3  but  less  so  for  map  B2  and  definitely 
inappropriate  in  case  of  map  CI.  Conversely,  the  CMSV  is  suitable  for  mapping 
CI,  but  it  is  not  appropriate  for  mapping  B2  and  even  less  so  for  mapping  A3. 
The  most  important  observation  from  figure  2  is  that  the  MMSV  is  indeed 
capable  of  an  adequate  mapping  in  all  three  cases.  It  is  interesting  to  observe 
that  the  MMSV  mimics  the  DMSV  in  case  of  a  'discrete  reality'  and  the  CMSV 
in  case  of  a  'continuous'  reality. 


246 


A3  -  DMSV  A3  -  MMSV  A3  -  CMSV 


B2  -  DMSV  B2  -  MMSV  B2  -  CMSV 


Figure  2. -Mapping  three  simulated  realities  (maps  A3,  B2  and  CI)  from  200 
observations  using  the  DMSV,  MMSV  and  CMSV. 

Validation 

In  order  to  evaluate  the  three  prediction  methods  the  mean  error  (ME),  root 
mean  square  error  (RMSE)  and  standardized  root  mean  square  error  (SRMSE) 
were  computed  for  each  of  the  9  realities.  These  statistics  were  computed  from 
all  remaining  points  in  the  map.  The  results  are  given  in  figure  3. 

The  mean  error  is  in  all  cases  quite  small.  This  is  not  surprising,  because 
unbiasedness  conditions  are  included  in  all  three  mapping  procedures. 
Differences  between  the  three  mapping  procedures  are  also  negligible. 

The  SRMSE  values  are  on  average  somewhat  larger  than  one,  particularly 
when  the  CMSV  is  applied  to  situations  in  which  the  soil  map  influence  is 
dominant.  This  may  be  caused  by  forcing  the  CMSV  upon  a  non-stationary 


247 


MMSV  or  DMSV  reality,  and  perhaps  also  because  the  kriging  variance  does 
not  include  the  uncertainty  in  estimating  the  variogram  (Christensen  1991). 

Most  interesting  are  the  RMSE  results.  In  this  case  we  do  see  meaningful 
differences  between  the  three  mapping  procedures.  The  results  demonstrate  that 
the  CMSV  is  inappropriate  for  the  A-maps,  whereas  the  DMSV  is  inappropriate 
for  the  C-maps.  An  exception  is  the  pure  nugget  map  C3,  where  all  mapping 
procedures  are  equally  good  (bad).  The  results  also  confirm  that  the  MMSV  is 
superior  for  the  B-maps.  Note  that  comparison  of  RMSE  values  between  maps  is 
difficult  here  because  these  are  affected  by  differences  in  spatial  autocorrelation. 


2.0 
1.5 
1.0 
0.5 
0.0 
-0.5 


Mean  Error 


D 


M  DMSV 


MMSV 


□  CMSV 


A1      A2  A3 


B1 


B2  B3 


C1      C2  C3 


20 
15 
10 
5 
0 


Root  Mean  Square  Error 


E3  DMSV 


-  HMMSV 


□  CMSV 


A1      A2  A3 


B1      B2  B3 


C1      C2  C3 


1.50 
1.25 


0.00 


Standardized  Root  Mean  Square  Error 


EDMSV 


MMSV 


□  CMSV 


A1      A2  A3 


B1      B2  B3 


C1      C2  C3 


Figure  3. -Validation  results  for  the  three  prediction  methods. 


248 


DISCUSSION  AND  CONCLUSIONS 


The  application  to  simulated  'realities'  shows  that  the  MMSV  interpolates 
well  on  the  entire  range  of  spatial  variation.  In  all  cases  is  the  MMSV  at  least  as 
good  as  the  DMSV  and  the  CMSV,  and  it  is  superior  in  situations  where  there 
is  abrupt  and  gradual  spatial  variation. 

The  simulation  exercise  also  shows  that,  depending  on  the  situation,  the 
DMSV  may  perform  much  worse  than  the  CMSV,  and  vice  versa.  This  means 
that  if  a  choice  between  these  two  models  is  to  be  made  (and  these  are  the  two 
models  most  often  used  in  practice),  then  it  must  be  taken  with  care.  This  may 
seem  obvious,  but  in  practice  the  choice  of  model  is  often  dominated  by 
irrelevant  factors,  such  as  background  and  experience  of  the  user.  And  even 
when  care  is  taken,  it  may  not  always  be  easy  to  decide  beforehand  whether 
abrupt  or  gradual  spatial  variation  prevails.  Therefore  the  flexibility  of  the 
MMSV  demonstrated  here  is  of  clear  importance,  because  it  implies  that  by 
adopting  the  MMSV  one  can  protect  oneself  against  using  the  wrong  model.  It  is 
as  if  one  can  leave  the  choice  between  abrupt  and  gradual  spatial  variation  to  the 
MMSV. 

The  MMSV  is  especially  advantageous  in  situations  where  abrupt  and 
gradual  spatial  variation  are  both  present.  Some  indication  of  whether  this  is  the 
case  can  be  obtained  from  the  intra-class  correlation  (Webster  and  Oliver  1990), 
but  more  informative  is  the  comparison  of  the  variograms  of  the  original 
attribute  and  its  residual.  A  mixed  form  of  spatial  variation  yields  a  residual 
variogram  that  is  substantially  lower  than  the  original  variogram.  Since  both 
variograms  can  easily  be  computed  one  can  thus  quickly  decide  whether  the 
MMSV  is  superior  to  the  CMSV  and  DMSV  for  a  given  situation. 

Another  approach  to  handling  abrupt  and  gradual  spatial  variation  both 
present  in  the  same  area  is  to  adopt  the  CMSV  separately  per  mapping  unit 
(Stein  et  al.  1988,  Voltz  and  Webster  1990).  The  main  difference  with  the 
MMSV  is  that  this  approach  excludes  the  presence  of  spatial  autocorrelation 
across  mapping  unit  boundaries.  Thus  it  is  likely  to  create  boundaries  even  when 
they  are  not  really  there.  As  mentioned,  the  MMSV  does  not  suffer  from  this 
problem  because  it  mimics  the  CMSV  under  such  circumstances.  We  consider  it 
a  major  advantage  of  the  MMSV  that,  although  it  is  meant  for  situations  in 
which  gradual  and  abrupt  spatial  variation  are  both  present,  it  will  also  perform 
well  when  spatial  variation  is  exclusively  gradual  or  exclusively  abrupt. 

REFERENCES 

Burgess,  T.M.  and  R.  Webster.  1980.  Optimal  interpolation  and  isarithmic 
mapping  of  soil  properties.  I.  The  semi-variogram  and  punctual  kriging.  J. 
Soil  Science  31,  315-331. 

Burrough,  P. A.  1993.  Soil  Variability:  a  late  20th  century  view.  Soils  and 
Fertilizers  56,  529-562. 


249 


Campbell,  D.J.,  D.G.  Kinninburgh  and  P.H.T.  Beckett.  1989.  The  soil  solution 

chemistry  of  some  Oxfordshire  soils:  temporal  and  spatial  variability.  J.  Soil 

Science  40,  321-339. 
Christensen,  R.  1991.  Linear  models  for  multivariate,  time  series,  and  spatial 

data.  New  York:  Springer. 
Cressie,  N.  1991.  Statistics  for  spatial  data.  New  York:  Wiley. 
Deutsch,  C.V.  and  A.G.  Journel.  1992.  GSLIB:  geostatistical  software  library 

and  user's  guide.  New  York:  Oxford  University  Press. 
Goovaerts,  P.  and  A.G.  Journel.  1995.  Integrating  soil  map  information  in 

modelling  the  spatial  variation  in  continuous  soil  properties.  European  J. 

Soil  Science  46,  397-414. 
Heuvelink,  G.B.M.  1996.  Identification  of  field  attribute  error  under  different 

models  of  spatial  variation.  Int.  J.  GIS  (in  press). 
Heuvelink,  G.B.M.  and  M.F.P.  Bierkens.  1992.  Combining  soil  maps  with 

interpolations  from  point  observations  to  predict  quantitative  soil  properties. 

Geoderma  55,  1-15. 

Kitanidis,    P.K.    1994.    Generalized   covariance    functions    in  estimation. 

Mathematical  Geology  25,  525-540. 
Nettleton,  W.D.,  B.R.  Brasher  and  G.  Borst.  1991.  The  taxadjunct  problem. 

Soil  Sci.  Soc.  Am.  J.  55,  421-427. 
Nortcliff,  S.  1978.  Soil  variability  and  reconnaissance  soil  mapping:  a  statistical 

study  in  Norfolk.  J.  Soil  Science  29,  403-418. 
Rogowski,  A.S.  and  J.K.  Wolf.  1994.  Incorporating  variability  into  soil  map 

unit  delineations.  Soil  Sci.  Soc.  Am.  J.  58,  403-418. 
Stein,  A.,  M.  Hoogerwerf  and  J.  Bouma.  1988.  Use  of  soil-map  delineations  to 

improve  (co-)  kriging  of  point  data  on  moisture  deficits.  Geoderma  43, 

163-177. 

Voltz,  M.  and  R.  Webster.  1990.  A  comparison  of  kriging,  cubic  splines  and 

classification  for  predicting  soil  properties  from  sample  information.  J.  Soil 

Science  31,  505-524. 
Webster,  R.  1994.  The  development  of  pedometrics.  Geoderma  62,  1-15. 
Webster,  R.  and  H.E.  De  La  Cuanalo.  1975.  Soil  transect  correlograms  of 

North  Oxfordshire  and  their  interpretation.  J.  Soil  Science  26,  176-194. 
Webster,  R.  and  M.A.  Oliver.  1990.  Statistical  methods  in  soil  and  land 

resource  survey.  Oxford:  Oxford  University  Press. 

BIOGRAPHICAL  SKETCH 

Gerard  B.M.  Heuvelink  is  a  geostatistician  with  the  Landscape  and 
Environmental  Research  Group,  Faculty  of  Environmental  Sciences,  University 
of  Amsterdam.  He  holds  an  M.S.  in  Applied  Mathematics  from  Twente 
Technical  University  and  a  Ph.D.  in  Geography  from  Utrecht  University. 

Johan  A.  Huisman  is  a  graduate  student  in  Physical  Geography  at  the 
Faculty  of  Environmental  Sciences,  University  of  Amsterdam. 


250 


Stratified  Kriging 
Using  Vague  Transition  Zones 

G.  BOUCNEAU1 ,  M.  VAN  MEIRVENNE1, 
O.  THAS2  and  G.  HOFMAN1 


Abstract.—  Stratification  followed  by  within-stratum  interpollation  is  a 
widespread  spatial  prediction  procedure  in  soil  inventories. 
Conventional  stratification  assumes  that  strata  are  delimited  by  accurate, 
crisp  boundaries.  However,  map  delineations  are  not  equally  accurate. 
Inaccuracy  can  result  from  the  inherent  gradual  nature  of  the  soil 
property,  or  errors  made  during  the  mapping  process.  This  paper 
presents  a  model  to  describe  the  vagueness  of  stratum  boundaries  and 
proposes  a  modified  within-stratum  kriging  algorithm  to  account  for  the 
uncertainty  due  to  this  vagueness.  The  procedure  was  used  to  map 
topsoil  sand  in  Belgium. 


INTRODUCTION 

Spatial  prediction  of  soil  properties  is  unavoidable  when  creating  a  soil 
information  system  (SIS).  However,  every  prediction  procedure  has  its 
shortcomings.  Choropleth  soil  maps  assume  abrupt  changes  at  their  boundaries 
and  interpolation  techniques,  like  kriging,  take  the  condition  that  variation  is 
gradual  and  stationary.  However,  in  reality  heterogeneous  areas  often  contain  both 
gradual  and  discontinuous  changes  of  soil  properties.  Therefore,  a  combination  of 
both  procedures  has  been  suggested  to  combine  the  best  of  both.  Stratification  of  a 
study  area  based  on  soil  map  delineations,  followed  by  within-stratum 
interpolation,  has  been  the  most  widespread  combined  approach.  Stein  et  al. 
(1988)  calculated  a  variogram  for  each  soil  type,  Voltz  &  Webster  (1990) 
computed  a  pooled  variogram  and  Van  Meirvenne  et  al.  (1994)  standardised  the 
pooled  variogram.  Other  approaches  have  been  the  formulation  of  a  mixed  model 
of  spatial  variation  (Heuvelink  &  Bierkens  1992),  or  more  recently  the 
introduction  of  a  non-parametric  indicator  approach  combined  with  qualitative 
information  provided  by  soil  or  geological  maps  (Goovaerts  &  Journel  1995). 


Soil  scientist.  Dept.  of  Soil  Management  and  Soil  Care.  University  of  Gent.  Coupure  653.  9000  Gent.  Belgium. 
'  Statistician.  Dept.  of  Applied  Mathematics.  Biometrics  and  Process  Control,  University  of  Gent.  Coupure  653.  9000 
Gent.  Belgium. 


251 


Conventional  stratification  assumes  strata  are  delimited  by  accurate,  crisp 
boundaries.  Though,  map  delineations  are  not  equally  accurate.  Inaccuracy  can 
result  from  the  inherent  fuzzy  nature  of  the  soil  property,  the  poor  field  work,  the 
distortion  or  shrinkage  of  the  original  paper  base  map  or  through  poor  quality 
digitisation.  Attempts  have  been  made  to  characterise  the  accuracy  of  soil  map 
boundaries.  Mark  &  Csillag  (1989)  conceptually  modified  the  cartographic 
epsilon  model  to  describe  the  uncertainty  of  class  membership  near  a  map 
boundary.  However,  they  do  not  elaborate  on  this  in  a  mathematical  way. 
Burrough  (1989)  mapped  the  transition  between  soil  strata  based  on  concepts  of 
the  fuzzy  set  theory.  Though,  he  did  not  make  any  link  with  spatial  interpolation. 

This  paper  aims  to  model  the  vagueness  of  stratum  boundaries  in  a 
mathematical  way  and  to  modify  the  within-stratum  kriging  algorithm  to  account 
for  the  uncertainty  due  to  this  vagueness. 

THEORY 

The  Vagueness  of  Stratum  Boundaries 

The  vagueness  of  a  stratum  boundary  is  described  by  examining  point 
observations  of  the  considered  soil  property  perpendicular  to  that  delineation 
(figure  1).  The  shortest  distance  between  observation  /'  of  property  y  in  stratum  j 
and  the  map  delineation  between  strata  A  and  B  is  %.  To  avoid  artefacts  at 
complex  combinations  of  boundaries,  each  observation  should  be  used  only  once, 

Stratum  delineation 


Figure  1.—  Examining  the  vagueness  of  a  stratum  boundary  using  point  observations  (left), 
the  sigmoidal  cosine  function  describing  the  vagueness  of  a  stratum  boundary 
(right). 


252 


i.e.  in  respect  to  the  nearest  delineation.  Then,  is  plotted  against  xtJ.  Distances  to 
observations  located  in  stratum  A  are  given  a  negative  sign,  those  within  stratum 
B  a  positive  one.  The  position  of  the  stratum  delineation  is  at  x  =  0.  The  following 
sigmoid  cosine  function  can  be  fitted  to  model  the  behaviour  of  y  at  both  sides  of 
the  boundary  (figure  2)  : 

y<j =  yA 


for 


for  Xj<x,,<x, 


for  Xy  >xL 


yy  =yA  +(yB-yA) 
y„  =yB 


l  -  cos' 


V 


[\xB  -xAJ 


[1] 


with  yA  and  yB  the  typical  value  of  the  soil  property  near  the  centre  of  stratum  A 
respectively  B  and  taM^I  the  width  of  the  transition  zone  between  both  strata 
This  function  was  chosen  since  it  allows  to  model  the  behaviour  of  a  soil  property 
near  a  soil  boundary  asymmetrically  in  respect  to  the  position  of  the  boundary. 


Within-stratum  Variograms 


For  each  stratum  the  within-stratum  variogram  was  calculated.  Observations 
located  within  narrow  transition  zones  of  inaccurately  mapped  boundaries  were 
excluded  from  variogram  calculations.  For  inaccurately  delimited  strata  this 
resulted  in  a  drastic  decrease  of  the  semivariance. 


Prediction  With  Vague  Boundaries 


In  conventional  stratification  adjacent  strata  are  delimited  by  an  infinitely  thin 
line  and  every  location  belongs  to  only  one  stratum.  Within-stratum  kriging  will 
then  result  in  only  one  kriging  estimate,  and  one  associated  variance,  for  every 
location  of  that  stratum.  In  our  procedure,  boundaries  represent  a  transition  zone, 
wide.  Within  this  zone  two  kriging  procedures,  K  =  a  and  K  =  b,  based 
on  the  variograms  and  data  of  adjacent  strata  A  respectively  B,  provide  two 

prediction  estimates  ya  and  yb  and  two  associated  estimates  of  the  variances  sa 

and  sb .  In  the  transition  zone  and  close  to  stratum  A  the  structure  of  the  variability 
will  be  more  similar  to  the  variogram  of  stratum  A  than  of  stratum  B  and  the 
validity  of  the  kriging  procedure  K  =  a  is  expected  to  be  higher  than  the  validity 
of  the  kriging  procedure  K  =  b.  The  reverse  holds  for  a  location  within  the 
transition  zone  and  close  to  stratum  B.  The  probabilities  P(K  =  a)  and  P(K  =  b) 
express  the  validity  of  kriging  procedures  K  =  a  respectively  K  =  b.  These 
probabilities  are  related  to  the  behaviour  of  the  examined  soil  property  in  the 
transition  zone.  So  Eq.  1  is  modified  as  follows: 


253 


for  x<xA  : 
for  xA<x<xB 


P(K  =  a)=  1  and  P(K  =  b)  =  0 


(( 


P(K  =  a)  =  l-cos 

(f 


x-  X 


B 


P(K  =  b)  =  cos 


x-x 


71 


B 


and 


[2] 


for  x  >xB: 


\\xA  -xBJ 


P(K  =  a)  =  OmdP(K  =  b)  =  7 


Outside  the  transition  zone  and  inside  stratum  A  is  P(K  =  a)  =  1  and  inside 
stratum  B  is  P(K  =  a)  =  0.  Inside  the  transition  zone  P(K  =  a)  decreases  from  1 
close  to  stratum  A  to  0  close  to  stratum  B.    The  reverse  holds  for  P(K  =  b). 

Estimators  of  the  unconditional  expected  value  E(Y)  and  its  variance  S2  are  given 
by  (see  appendix  for  the  derivations)  : 


E(Y)=YaP(K  =  a)  +  YbP(K  =  b) 
S2  =  P(K  =  a)P(K  =  b)(Ya  -Y„)2- 

P(K  =  a)P(K  =  b)(Sa  -  S„ )2  +S2a  P(K  =  a)  +  SfP(K  =  b) 


[3] 
[4] 


To  illustrate  the  influence  of  P(K  =  a)  and  P(K  =  b)  on  Sr,  three  situations  were 
distinguished  : 

1.  Ya  is  twice  as  large  as  Yb  and      and  S%  are  equal.  This  results  in  a  quadratic 
curve  path  of  the  variance  along  the  transition  zone  (figure  2a)  with  a 

maximum  at  P(K  =  a)  =  P(K  =  b)  =  0.5.  Here  S1  is  3.25  times  S2a. 

2.  Ya  and  Yb  are  equal  and  Sa  is  twice  as  large  as  Sb  .  This  results  in  a  linearly 

increasing  estimation  variances  between  Sa  and  Sb  (figure  2b). 

3.  In  practice,  mostly  both  the  prediction  estimates  and  their  variances  will 

differ.  In  figure  2c  both  Ya  is  twice  as  large  as  Yb  and  Sa  is  twice  as  large  as 

Sb  .  The  result  is  a  combination  of  both  preceding  situations,  with  a  maximum 

of  S1  being  3.71  times  as  large  as  S2a.  It  is  clear  that  the  difference  in 
prediction  estimates  dominates  the  difference  in  variances. 

(a)  (b)  (c) 


P(K=a)\.0       0.5        0.0        1.0       0.5       0.0         1.0  0.5 
P(K=b)  0.0       0.5        1.0        0.0       0.5        1.0        0.0       0.5  1.0 

Figure  2  —  The  prediction  variance  in  the  transition  zone  :  (a):  Ya  =  2  Yb  and  S  *  =  S  2b  ; 
(b):  Ya  =  Yb  and  S2a  =2S2b  ;  (c):  Ya  =  2  Yb  and  S2  =  2S2b. 


254 


It  makes  sense  that  the  prediction  variance  inflates  at  the  position  of  a  retained 
map  boundary  since  this  boundary  itself  represents  uncertain  information.  If  one 
chooses  to  make  use  of  this  information,  its  uncertainty  should  be  added  to  the 
prediction  variance. 


The  procedure  was  tested  on  the  top  soil  sand  fraction  of  West-Flanders,  Belgium. 
More  details  can  be  found  in  Boucneau  et  al.  (1996).  Stratum  boundaries  were 
characterised  using  Eq.  1.  Figure  3  presents  an  accurately  mapped  boundary 
between  two  strata  B  and  C  (Strata  are  indicated  in  figure  4a),  having  a  narrow 
transition  zone  of  500  m  (left),  and  an  inaccurately  mapped  boundary  between  two 
strata  A  and  C,  having  a  wide  transition  zone  of  1500  m  (right). 

A  prediction  estimate  of  the  topsoil  sand  and  its  associated  variance  was 
produced  based  on  Eqs.  3  and  4.  Figure  4  gives  a  detail  of  the  map  results  and 
allows  to  compare  the  proposed  procedure  with  the  conventional  stratification 
procedure  (i.e.  without  a  transition  zone).  Three  strata  can  be  distinguished:  two 
sandy  strata  in  the  upper  left  corner  (stratum  A)  and  the  lower  right  corner 
(stratum  B)  and  a  stratum  with  low  sand  content  (stratum  C)  between  both.  Within 
the  wide  transition  zone,  between  the  strata  A  and  C,  the  proposed  procedure 
recognises  the  prediction  estimates  are  uncertain  by  attributing  them  a  high 
prediction  variance  (figure  4c).  Between  strata  B  and  C  the  boundary  is  accurately 
mapped.  Here  the  transition  zone  is  narrow  and  high  prediction  variances  are 
absent. 


RESULTS 


Stratum  B 
•  Stratum  C 


Stratum^ 
•  Stratum  C 


100-, 


T3 

in 


40- 


4 


20- 


'  Transition 


zone 


Transition 


zone^ 


u~1 — 1 — I — 1    I  1 — l — 1 — l 
-2000-1000    0     1000  2000 


2000-1000    0     1000  2000 


Distance  to  the  boundary  (m) 


Distance  to  the  boundary  (m) 


Figure  3.  —  Narrow  (left)  and  wide  (right)  transition  zones  between  two  strata. 


255 


(c) 


(d) 


Prediction  estimate  (%) 

Not  Mapped 

0-  15 
15  -  30 


Prediction  variance  (%  ) 

Not  Mapped 

0  -  50 
50  -  100 
100  -  150 
150  -  200 
200  -  250 
250  -  300 
300  -400 

Figure  4.—  Detail  of  the  topsoil  sand  map:  the  prediction  estimate  based  on  the  proposed 
(a)  and  the  conventional  procedure  (b);  the  prediction  variance  based  on  the 
proposed  (c)  and  the  conventional  procedure  (d). 

Only  observations  outside  the  transition  zone  are  used  for  computing  the 
variogram  (figure  5).  As  a  result,  the  prediction  variance  in  strata  A  and  C  and 
outside  the  transition  zones,  is  considerably  lower  in  the  proposed  procedure  than 
in  the  conventional  procedure  (figure  4c  and  4d). 


o 


ON 


Q 


Using  all  observations 

Observations  in  transition 
zones  of  inaccurately  mapped 
boundaries  omitted 


— l — 1 — I 

1000  2000  3000  4000 
Lag(m) 

Figure  5.—  The  within-stratum  semivariogram  of  stratum  C. 


256 


CONCLUSIONS 


In  conventional  stratification  the  study  area  is  sharply  partitioned  in  several 
strata  delimited  by  a  thin  line.  As  a  result,  the  within-stratum  variogram  is  based 
on  all  data  within  each  stratum.  However,  soil  attribute  values  of  locations 
situated  near  the  border  of  a  stratum  often  differ  from  the  typical  value  of  that 
attribute.  These  observations  can  increase  the  semivariance  inside  a  stratum 
considerably.  In  this  way  the  variability  inside  the  transition  zone  influences  the 
estimate  of  the  variance  within  the  entire  stratum.  Our  procedure  overcomes  this 
problem  by  characterising  the  transition  zone  between  strata.  Only  observations 
outside  the  transition  zone  are  used  for  computing  the  variogram.  For  each 
location  in  a  transition  zone  the  probabilities  P(K  =  a)  and  P(K  =  b)  are  quantified 
and  incorporated  in  the  prediction  estimate  and  prediction  variance.  In  this  way, 
the  variance  due  to  the  vagueness  of  stratum  boundaries  is  concentrated  near  these 
boundaries. 

REFERENCES 

Boucneau,  G.,  Van  Meirvenne,  M.,  Thas,  O.  &  Hofrnan  G.  1996.  Stratified  kriging 

using  selected  soil  map  boundaries  modelled  as  vague  transition  zones. 

Submitted  to  European  Journal  of  Soil  Science. 
Burrough,  P. A.  1989.  Fuzzy  mathematical  methods  for  soil  survey  and  land 

evaluation.  Journal  of  Soil  Science,  40:  477-492. 
Goovaerts.  P.  &  Journel.  A.G.  1995.  Integrating  soil  map  information  in 

modelling  the  spatial  variation  of  continuous  soil  properties.  European  Journal 

of  Soil  Science.  46:  397-414. 
Heuvelink.  G.B.M.  &  Bierkens,  M.F.P.  1992.  Combining  soil  maps  with 

interpolations  from  point  observations  to  predict  quantitative  soil  properties. 

Geoderma,  55:  445-468. 
Mark,  D.M.  &  Csillag.  F.  1989.  The  nature  of  boundaries  on  "Area-Class"  maps. 

Cartographica,  26:  65-78. 
Stein,  A.,  Hoogerwerf,  M.  &  Bouma.  J.  1988.  Use  of  soil-map  delineations  to 

improve  (co-)kriging  of  point  data  on  moisture  deficits.  Geoderma,  43:  163- 

177. 

Van  Meirvenne,   M.,   Scheldeman,   K.,   Baert,   G.   &   Hofman,   G.  1994. 

Quantification  of  soil  textural  fractions  of  Bas-Zai're  using  soil  map  polygons 

and/or  point  observations.  Geoderma,  62:  69-82. 
Voltz.  M.  &  Webster.  R.  1990.  A  comparison  of  kriging,  cubic  splines  and 

classification  for  predicting  soil  properties  from  sample  information  Journal  of 

Soil  Science  41:473-490. 


257 


APPENDIX 


Prediction  estimates  are  obtained  based  on  two  kriging  procedures  K=a  and  K=b 
which  are  based  on  variograms  of  stratum  A  respectively  B.  The  expected  value  of 
the  examined  variable  Y  if  the  kriging  procedure  K=k  is  valid  is  yk  =  E(Y\ K=k). 
The  unconditional  expected  value  y  is  then  obtained  by  summation  over  the  set  of 
procedures  K={a,b}: 

y  =  E(Y)=  Z  E(Y\K  =  k)P(K  =  k)  =  yaP(K  =  a)  +  ybP(K  =  b) 

keK 

2  2  2  2 

The  variance  a  is  defined  by:  a  =Var(Y)  =E(Y  )-{E(Y)}  .Working  out  each  term: 
E(Y)2  =  y2aP2  (K=a)  +  y2bP2  (K  =  b)  +  2yayb  P(K  =  a)P(K  =  b) 

E(Y2)  =  I  E(Y2\K=k)P(K=k)  =  (y2+o2)P(K  =  a)  +  (y2+G2)P(K  =  b) 

ksK 

Since: 

Var(Yk \K  =  k)  =  E(Y2 \K  =  k)-  (E(Yk \K  =  kjf 

E(Y2\K  =  k)  =  Var(Yk \K  =  k)  +  (E(Yk \K  =  kjj  =  y\  +g2 
And  substituting  them,  the  variance  a  is  given  by: 

o  2  =  (yl  +a2)P(K  =  a)  +  (y2  +v2b)P(K  =  b)-  y2  P2(K  =  a)- 
y2P2(K  =  b)-2  yayb  P(K  =  a)P(  K  =  b) 
=  y2P(K  =  a)(l - P(K  =  a))  +  y2bP(K  =  b)(l -  P(K  =  b)) - 
2yaybP(K  =  a)P(K  =  b)  +o2P(K  =  a)  +  a2P(K  =  b) 

=  P(K  =  a)P(K  =  b)(ya-yb)2+  o2  P(K  =  a)  +  a2  P(K  =  b) 
The  unbiased  estimator  for  the  unconditional  mean  is: 

E(Y)  =  YaP(K  =  a)  +  YbP(K  =  b) 
Since:  y  =  ya  P(K  =  a)  +  yb  P(K  =  b) 
The  unbiased  estimator  for  the  unconditional  variance  is: 

S2  =  P(K  =  a)P(K  =  b)(Ya  -Yb)2-  P(K  =  a)P(K  =  b)(Sa  -Sb)2  + 

S2aP(K  =  a)  +  S2bP(K  =  b) 
Since:  a2  =  P(K  =  a)P(K  =  b)(ya  - yb)2  +<j2 P(K  =  a)  +  ob2 P(K  =  b) 

BIOGRAPHICAL  SKETCH 

Geert  Boucneau  prepares  a  Ph.D.  in  Soil  Information  Systems.  Marc  Van 
Meirvenne  is  professor  Soil  Information  Systems  and  Spatial  Data  Analysis. 
Olivier  Thas  is  active  in  the  domain  of  Environmental  Statistics.  Georges  Hofman 
is  professor  Soil  Science  and  Soil  Fertility. 


258 


The  Effect  of  Uncertain  Locations  on 
Disease  Cluster  Statistics 

Geoffrey  M.  Jacquez1  and  Lance  A.  Waller2 

Abstract.  Disease  clusters  are  space-time  aggregations  of  adverse  health 
events  such  as  birth  defects  and  cancer.  Because  of  incomplete 
information  the  exact  locations  of  health  events  are  often  unknown,  but 
to  date  the  effect  of  uncertain  locations  on  disease  cluster  statistics  has 
not  been  assessed.  This  research  demonstrates  that  uncertain  locations,  if 
not  accounted  for,  lead  to  test  statistics,  null  distributions  and  p-values 
that  differ  substantially  from  those  based  on  actual  locations,  and 
increases  the  probability  of  type  II  error  (false  negatives).  Further 
research  is  needed  to  develop  statistical  methods  and  software  to  account 
for  uncertain  locations  in  disease  cluster  investigations.  

INTRODUCTION 

Public  health  agencies  use  cluster  statistics  to  assess  whether  an  alleged 
aggregation  of  disease  cases  is  statistically  unusual.  The  health  agency  may 
choose  to  stop  a  cluster  investigation  when  the  alleged  cluster  appears  to  be  due  to 
chance,  otherwise  a  more  detailed  and  expensive  investigation  may  be  warranted 
(Centers  for  Disease  Control  1990).  Point-based  statistics  assume  exact 
geographic  locations  and  are  often  used  when  health  events  are  rare  and  the 
number  of  observations  is  small.  They  are  coming  into  increasing  use  as  spatially 
referenced  health  data  become  commonly  available  (Openshaw  1991),  and  models 
of  spatial  point  processes  (Diggle  1993,  Lawson  and  Waller  1996),  become 
increasingly  sophisticated.  However,  the  assumption  of  exact  locations  upon 
which  point-based  statistics  are  founded  is  often  invalid.  Does  the  violation  of 
this  assumption  make  a  difference? 

Sources  Of  Location  Uncertainty 

Point  based  statistics  assume  data  of  the  form  (x,  y,  t)  where  x,  y  is  the 
geographic  coordinate  (such  as  place  of  residence)  and  t  is  the  date  of  onset, 
diagnosis  or  death.  More  often  than  not  this  location  model  is  inappropriate 
because  locations  of  the  health  events  are  uncertain.  This  uncertainty  has  several 
sources.  It  emerges  when  centers  of  areas  such  as  zip-code  zones  or  census  tracts 
are  used  instead  of  exact  place  of  residence.  Uncertainty  also  arises  when  data  are 
gridded  (as,  for  example,  in  raster-based  GIS)  and  the  coordinates  of  the  nearest 
grid  node  are  used  instead  of  exact  locations.  Conceptually,  locations  are  almost 
always  uncertain  because  humans  are  mobile  rather  than  sessile  and  because 

'President,  BioMedware,  Ann  Arbor,  Ml. 

: Assistant  Professor,  University  of  Minnesota,  Minneapolis 


259 


health  events  and  their  causative  exposures  may  occur  anywhere  within  a  persons 
activity  space.  Tobler  et  al.  (1995)  observed  that  in  modern  society  a  persons 
daily  activity  space  is  approximately  1 5  km  and  varies  widely.  Exact  locations  do 
not  represent  this  location  uncertainty. 

There  are  many  examples  of  uncertain  locations  in  the  literature.  Cuzick  and 
Edwards  (1990)  used  the  centers  of  postal  code  zones  to  represent  place  of 
residence  in  their  study  of  childhood  leukemia  in  North  Humberside.  Waller  et  al. 
(1995)  mapped  locations  of  cases  of  childhood  leukemia  in  Sweden  at  parish 
centroids,  and  evaluated  proximity  to  nuclear  power  plants  using  several  methods. 
Location  uncertainty  arises  in  these  two  examples  because  area  centroids  are  used 
instead  of  exact  locations.  In  a  study  notable  for  its  spatial  resolution,  Lawson  and 
Williams  (1994)  used  place  of  residence  to  assess  a  possible  cluster  of  deaths 
from  respiratory  cancer  near  a  smelter  in  Armadale,  Scotland.  Indoor  air  quality 
is  known  to  differ  substantially  from  that  outdoors,  and  an  unknown  proportion  of 
each  individuals  exposure  to  smelter  fumes  presumably  occurred  outside  of  the 
home.  In  this  example  location  uncertainty  arises  because  place  of  residence  is 
used  to  represent  exposures  that  occurred  throughout  each  persons  activity  space. 

We  have  not  evaluated  whether  or  not  uncertain  locations  had  an  effect,  if  any, 
on  these  studies.  Rather,  our  objective  in  citing  these  examples  is  to  give  readers 
a  feeling  for  the  ubiquity  of  the  location  uncertainty  problem.  In  practice 
geographic  locations  are,  to  a  greater  or  lesser  extent,  inherently  uncertain  because 
they  are  used  as  proxy  measures  of  exposure.  If  we  had  more  information  we 
could  explore  dose-response  relationships  using  controlled  studies.  When  such 
detailed  knowledge  is  lacking  we  instead  use  spatial  relationships  among  cases, 
and  their  proximity  to  putative  sources  of  hazard,  as  an  uncertain  measure  of 
exposure.  However,  almost  all  point-based  cluster  statistics  assume  the  spatial 
coordinates  of  health  events  are  precise.  What  is  the  impact,  if  any,  of  using 
statistics  that  assume  exact  locations  when  the  locations  are  actually  uncertain? 

METHODS 

Perhaps  the  most  frequently  used  paradigm  for  dealing  with  uncertain  locations 
is  the  centroid  model.  This  model  assigns  all  cases  occurring  within  an  area  the 
location  of  that  area's  centroid.  It  was  used  explicitly  in  the  studies  in  Humberside 
and  Sweden  and  is  implied  in  the  Armadale  study.  It  obtains  whenever  data  are 
gridded  and  is  common  in  environmental  epidemiology  where  insufficient 
knowledge  of  exposures  and  confidentiality  concerns  preclude  exact  knowledge  of 
health  event  locations.  Using  a  simulated  AIDS  epidemic  in  Michigan,  we 
explored  the  performance  of  three  disease  cluster  statistics  at  three  levels  of 
spatial  resolution.  We  compared  the  test  statistics,  null  distributions,  p-values  and 
statistical  power  under  the  centroid  model  to  those  obtained  when  actual  locations 
were  used. 


260 


Cluster  Statistics 

Three  statistics  for  space-time  clustering  were  compared,  Mantel  (1967),  Knox 
(1964)  and  k-NN  (Jacquez  1996a).  These  commonly  used  tests  were  chosen 
because  they  are  based  on  different  spatial  proximity  measures;  geographic 
distance,  adjacencies  and  nearest  neighbor  relationships,  respectively  (Jacquez 
1996b).  We  choose  tests  based  on  several  different  proximity  measures  so  that  we 
could  generalize  the  results  to  a  wider  range  of  other  cluster  tests. 

Mantel's  test  is  based  on  regression  of  the  space  distance  between  pairs  of  cases 
on  the  time  distance  between  case  pairs.  This  study  used  the  reciprocal 
transformation  suggested  by  Mantel.  The  Knox  test  employs  space  and  time 
critical  distances  to  classify  pairs  of  cases  as  either  near  or  far  in  space  and  time. 
For  example,  two  cases  are  considered  near  to  one  another  when  the  distance 
between  them  is  less  than  the  space  critical  distance.  In  this  study  we  selected 
critical  values  to  maximize  the  Knox  test  statistic,  and  used  these  critical  values 
for  all  of  the  simulations.  While  choosing  critical  values  in  this  fashion  precludes 
statistical  hypothesis  testing,  it  was  deemed  appropriate  for  the  purpose  of  the 
exploratory  analyses  in  this  study.  The  k-NN  statistic  is  the  intersection  of  the 
space  and  time  nearest  neighbor  relationships  and  is  described  by  Jacquez 
(1996a).  All  cluster  statistics  were  calculated  using  the  Stat!  software  (Jacquez 
1994). 


Figure  1.  Boundaries  corresponding  to  four  levels  of  geographic  uncertainty. 


Geography 

Four  levels  of  geographic  uncertainty  were  evaluated:  none,  county,  bicounty 
and  region  (Figure  1).  Places  of  residence  of  the  at-risk  population  are  shown  in 
la  by  the  symbol  and  were  sampled  based  on  population  density  distributions 
of  the  global  demography  project  (Tobler,  Deichmann  et  al.  1995).  Centroids  and 
areas  are  shown  for  counties  (lb),  bicounties  (lc)  and  regions  (Id).  Area 
centroids  are  shown  as  squares. 

HIV  Simulation 

A  simulated  epidemic  of  AIDS  in  Michigan  was  used  to  generate  space-time 
distributions  of  cases,  and  is  based  on  the  model  of  Jacquez  et  al  (1994).  This  HIV 


261 


model  was  chosen  because  it  describes  the  dynamics  of  the  epidemic  and  because 
the  transmission  parameters  have  been  estimated.  The  model  was  simplified  for 
this  research  and,  while  the  simplified  version  cannot  be  said  to  represent  HIV  in 
human  populations,  it  is  a  contagious  process  and  should  result  in  space-time 
clusters.  If  the  cluster  statistics  perform  poorly  for  this  model  we  have  little 
reason  to  expect  them  to  perform  well  for  diseases  such  as  environmentally 
caused  cancer,  for  which  space-time  clustering  should  be  substantially  weaker. 

The  modeled  disease  has  4  stages:  healthy,  primary  infection,  seropositive,  and 
deceased.  The  vector  of  transmission  probabilities  was  (0.0,  0.3,  0.001,  0.0).  Each 
element  of  this  vector  describes  the  probability,  for  a  healthy  individual,  of 
infection  per  contact  with  an  individual  in  disease  stage  i.  The  underlying 
population  consists  of  500  individuals  with  a  heterogeneous  geographic 
distribution  (Figure  la).  Two  of  these  were  selected  as  seeds  to  start  the 
epidemic.  At  each  time  step  each  individual  contacts  1  of  any  of  4  nearest 
neighbors,  and  a  transmission  event  can  occur  when  the  contacted  individual  is 
infectious.  Transition  from  healthy  to  primary  infection  does  not  occur  in  the 
absence  of  a  transmission  event.  Probabilities  describe  transition  from  one 
disease  stage  to  another.  In  each  round  there  is  a  5%  probability  that  individuals 
in  the  primary  infection  stage  will  become  seropositive,  while,  on  average,  1%  of 
seropositives  enter  the  terminal  stage. 

The  simulation  is  a  three-step  process;  first  transmission  events  (which  cause 
healthy  individuals  to  transit  to  primary  infection)  are  evaluated.  Next,  stage 
transitions  are  determined  using  transition  probabilities.  Finally,  after  each 
iteration  is  complete  the  space-time  location  of  each  new  seropositive  is  recorded 
and  the  simulation  halts  when  50  individuals  become  seropositive.  The 
seropositive  stage  thus  was  used  for  the  reporting  of  disease  clusters.  The  model 
is  stochastic  and  uses  a  uniform  random  number  generator  to  evaluate  the  discrete 
transitions  for  each  individual  in  the  population.  It  is  realistic  because 
seropositives  are  monitored,  and  because  stochasticity  is  introduced  through 
probabilistic  processes  of  contact,  transmission  and  state-to-state  transition. 

Simulation  Protocol 

The  HIV  simulation  model  was  run  252  times  using  the  actual  locations  shown 
in  Figure  la,  and  new  locations  of  the  2  initial  infectives  were  chosen  for  each 
run.  Each  of  these  runs  is  a  distinct  realization  of  the  AIDS  epidemic  and  is  called 
an  'AIDS  simulation'.  Each  AIDS  simulation  produced  a  list  of  coordinate  pairs 
describing  the  actual  locations  of  50  seropositives.  Centroid  locations  for  the 
county,  bicounty  and  regional  aggregation  levels  were  assigned  as  appropriate 
(Figure  1).  For  example,  at  the  county  level  a  case  occurring  in  county  /  was 
assigned  the  coordinates  of  that  county's  centroid.  The  actual  test  statistic  (r^) 
and  reference  (null)  distribution  (gA)  were  calculated  using  the  actual  locations, 
and  usually  are  not  observable  in  field  situations  when  centroids  are  used.  The 


262 


centroid  test  statistic  (Vc)  and  its  null  distribution  (gc)  were  then  calculated  using 
the  centroid  locations,  and  correspond  to  statistics  calculated  in  studies  where 
centroid  locations  are  used.  The  gA  and  gc  distributions  used  249  randomization 
runs  to  generate  reference  distributions  expected  under  the  null  hypothesis  of  no 
association  between  a  case's  location  and  the  time  at  which  the  case  became 
seropositive.  This  was  accomplished  by  reallocating  the  times  the  cases  became 
seropositive  at  random  across  the  case  locations.  These  statistics  and  distributions 
were  recorded  for  each  AIDS  simulation  and  for  each  cluster  statistic. 

In  all,  252  AIDS  simulations  were  run.  249  Monte  Carlo  randomizations  were 
used  to  generate  each  statistical  null  distribution.  188,244  Monte  Carlo 
randomizations  were  needed  to  generate  the  gA  distributions  (252  AIDS 
simulation  x  3  cluster  tests  x  249  randomizations)  and  564,732  Monte  Carlo 
randomizations  were  needed  to  generate  the  gc  distributions  (252  AIDS 
simulation  x  3  cluster  tests  x  3  uncertainty  levels  x  249  randomizations).  A  grand 
total  of  752,976  Monte  Carlo  randomizations  were  conducted  using  spatial 
uncertainty  modeling  software  being  developed  by  BioMedware. 

RESULTS 

Inquiry  focused  on  four  questions.  How  does  uncertainty  regarding  the  spatial 
locations  of  health  events  affect  (1)  the  test  statistic,  (2)  its  null  distribution,  (3) 
p-values,  and  (4)  statistical  power  of  point-based  cluster  statistics? 

Figure  2  shows  the  results  at  the  county  level.  For  Mantel's  test,  test  statistics 
calculated  from  the  centroid  locations  were  consistently  lower  than  those 
calculated  using  actual  locations,  and  the  upper  95%  critical  value  of  the  gc 
reference  distribution  was  not  highly  correlated  with  the  95%  critical  value  of  the 
gA  reference  distribution  calculated  from  the  actual  locations.  As  a  result,  p-values 
using  the  centroid  locations  were  essentially  independent  of  p-values  based  on  the 
actual  locations  (Figure  2a).  This  pattern  of  results  holds  for  both  the  Knox  test 
(Figure  2b)  and  the  A>NN  test  (Figure  2c).  Results,  not  shown  due  to  space 
constraints,  at  the  bicounty  and  regional  aggregation  levels  are  as  bad  or  worse 
and  demonstrate,  for  the  HIV  model  and  at  the  spatial  scales  considered,  that 
p-values  obtained  under  the  centroid  location  model  differ  substantially  from 
p-values  based  on  actual  locations.  This  is  consistent  with  the  results  of  Waller 
and  Turnbull  (1993),  who  reported  changes  in  p-values  at  the  census  tract  and 
block  group  aggregation  levels  for  focused  cluster  tests. 

For  the  3  cluster  statistics  considered  there  is  a  substantial  decrease  in  statistical 
power  at  the  county,  bicounty  and  regional  aggregation  levels.  Statistical  power 
was  quantified  as  the  proportion  of  simulation  runs  for  which  the  p-value  was  less 
than  or  equal  to  0.05.  Using  actual  locations  the  statistical  power  was  0.592, 
0.480  and  0.792  for  the  Mantel,  Knox  and  k-NN  tests,  respectively.  At  the  county 
level  the  drop  in  statistical  power  ranged  from  3.67%  (Knox)  to  36.62%  (£-NN). 
At  the  bicounty  level  the  drop  ranged  from  6.58%  (Knox)  to  44.1%>  (£-NN).  At 


263 


Figure  2.  Results  at  the  county  level,  centroid  p-values  vs.  actual  p-values  for  Mantel  (a) 
Knox  (b)  and  A-NN  (c)  tests. 

the  regional  level  the  drop  in  statistical  power  ranged  from  22.60%  (Knox)  to 
59.83%  (jfc-NN). 

These  results  illustrate  that  p-values  from  centroid  data  bear  little  resemblance 
to  p-values  based  on  actual  locations.  This  causes  a  loss  in  statistical  power  when 
using  centroid  locations,  and  an  increase  in  type  II  error  (false  negatives). 

CONCLUSION 

As  mentioned  in  the  introduction,  cluster  statistics  are  one  of  several  criteria 
used  by  public  health  agencies  to  determine  the  future  course  of  a  cluster 
investigation.  In  such  situations,  p-values  from  disease  cluster  tests  are  used  to 
quantify  how  unusual  an  observed  aggregation  of  cases  is.  A  very  small  p-value 
suggests  the  aggregation  was  unlikely  to  have  been  a  chance  event,  and  additional 
effort  may  be  needed  to  identify  environmental  exposures  that  might  explain  the 
cluster.  While  generalization  from  a  single  simulation  study  is  difficult,  our 
results  suggest  that  p-values  based  on  uncertain  locations  are  unreliable  and 
should  not  be  used  as  a  quantitative  basis  for  determining  the  future  course  of 
cluster  investigations. 

Several  authors  have  commented  on  the  apparent  lack  of  statistical  power  in 
cluster  investigations.  This  study  suggests  that  actual  clusters  may  be  missed 
because  of  the  loss  of  statistical  power  due  to  location  uncertainty. 

Even  when  place  of  residence  is  known  to  the  level  of  the  street  address, 
geographic  location  is  often  uncertain  because  health  events  occur  throughout  a 
person's  daily  activity  space,  and  because  exposures  (e.g.  transmission  events  for 
infectious  diseases,  exposure  to  radiation  and  carcinogens  for  cancer,  etc.) 
underlying  health  events  occur  outside  as  well  as  inside  the  home.  This  indicates 
a  substantial  need  for  ways  of  accounting  for  uncertain  locations  in  general,  and 
for  point-based  statistics  in  particular. 

There  are  broader  implications  because  the  cluster  tests  considered  are  similar  to 
statistics  used  in  other  fields.  For  example,  nearest  neighbor  statistics  are  used 
extensively  in  ecology,  and  Mantel's  test  can  be  expressed  as  the  Pearson  product 
moment  correlation  which  is  used  extensively  in  the  environmental  sciences.  All 
GIS's  employ  data  models  that,  depending  on  the  GIS  operation,  may  abstract 
locations  to  grid  nodes  or  polygon  centers,  thereby  introducing  location 


264 


uncertainty.  This  study  therefore  may  have  implications  for  statistical  analyses  of 
data  managed  by  Geographic  Information  Systems. 

In  conclusion,  the  software  and  methods  developed  in  this  research  allow  one  to 
assess,  given  a  specific  geography  and  hypothesized  space-time  process,  the 
impact  of  uncertain  locations  on  test  statistics,  reference  distributions,  p-values 
and  statistical  power.  Generalization  from  the  single  simulated  disease  process 
and  geographies  considered  is  problematic.  Will  the  pattern  of  results  hold  at  finer 
geographic  resolutions?  Because  of  their  use  in  small  area  studies,  it  seems 
particularly  important  to  assess  the  impact  of  uncertain  locations  for  census  tract 
and  zip-code  zone  geography.  Now  that  a  substantial  effect  at  the  county  level  has 
been  demonstrated,  the  methods  developed  in  this  preliminary  study  can  be  used 
to  address  this  question  and  others  like  it. 

ACKNOWLEDGMENTS 

This  research  was  funded  by  Small  Business  Innovation  Research  grant  R43 
CA65366-01A1  from  the  National  Cancer  Institute.  Its  contents  are  solely  the 
responsibility  of  the  authors  and  do  not  necessarily  represent  the  official  views  of 
the  NCI. 

REFERENCES 

Centers  for  Disease  Control  1990.  Guidelines  for  investigating  clusters  of  health 
events.  Morbidity  and  Mortality  Weekly  Report  39:  1-23. 

Cuzick,  J.  and  R.  Edwards  1990.  Spatial  clustering  for  inhomogeneous 
populations.  Journal  of  the  Royal  Statistical  Society  52:  73-104. 

Diggle,  P.  J.  1993.  Point  process  modelling  in  environmental  epidemiology. 
Statistics  for  the  environment.  V.  Barnett  and  K.  F.  Turkman.  New  York,  John 
Wiley  &  Sons  Ltd.  pp.  89-110. 

Jacquez.  G.  M.  1994.  Stat!  statistical  software  for  the  clustering  of  health  events. 
Ann  Arbor,  MI,  BioMedware. 

Jacquez,  G.  M.  1996a.  Disease  cluster  statistics  for  imprecise  space-time 
locations.  Statistics  in  Medicine  15:  In  press. 

Jacquez,  G.  M.  1996b.  A  ^-nearest  neighbor  test  for  space-time  interaction. 
Statistics  in  Medicine  In  press. 

Jacquez.  J.  A.,  J.  S.  Koopman,  et  al.  1994.  Role  of  the  primary  infection  in 
epidemics  of  HIV  infection  in  gay  cohorts.  Journal  of  Acquired  Immune 
Deficiency  Syndromes  7:  1169-1184. 

Knox,  G.  1964.  The  detection  of  space-time  interactions.  Applied  Statistics  13: 
25-29. 

Lawson,  A.  B.  and  L.  A.  Waller  1996.  A  review  of  point  pattern  methods  for 
spatial  modelling  of  events  around  sources  of  pollution.  Environmetrics  In  press. 


265 


Lawson,  A.  B.  and  F.  L.  R.  Williams  1994.  Armadale  :  A  case-study  in 
environmental  epidemiology.  Journal  of  the  Royal  Statistical  Society  157: 
285-298. 

Mantel,  N.  1967.  The  detection  of  disease  clustering  and  a  generalized  regression 
approach.  Cancer  Research  27:  201-218. 

Openshaw,  S.  1991.  A  new  approach  to  the  detection  and  validation  of  cancer 
clusters  :  A  review  of  opportunities,  progress  and  problems.  Statistics  in 
Medicine.  F.  Dunstan  and  J.  Pickles.  Oxford,  Clarendon  Press:  49-63. 

Tobler,  W.,  U.  Deichmann,  et  al.  1995.  The  global  demography  project.  Santa 
Barbara,  National  Center  for  Geographic  Information  and  Analysis. 

Waller,  L.  A.  and  B.  W.  Turnbull.  1993.  The  effects  of  scale  on  tests  of  disease 
clustering.  Statistics  in  Medicine  12:  1869-1884. 

Waller,  L.  A.,  B.  W.  Turnbull,  et  al.  1995.  Detection  and  assessment  of  clusters  of 
disease:  An  application  to  nuclear  power  plant  facilities  and  childhood  leukemia 
in  Sweden.  Statistics  in  Medicine  14:  3-16. 

BIOGRAPHICAL  SKETCH 

Geoffrey  M.  Jacquez  is  President  of  BioMedware,  Inc.  of  Ann  Arbor,  Michigan. 
He  holds  an  M.S.  in  resource  policy  and  law  from  the  University  of  Michigan,  and 
a  Ph.D.  (1989)  in  ecology  and  evolution  from  SUNY  Stony  Brook. 

Lance  Waller  is  Assistant  Professor  of  Biostatistics  in  the  School  of  Public 
Health  at  the  University  of  Minnesota.  He  received  his  Ph.  D.  in  operations 
research  from  Cornell  University  in  1991. 

BioMedware  was  founded  in  1990,  develops  software  and  statistical  methods 
for  the  environmental  and  health  sciences,  and  is  the  recipient  of  SBIR  and  STTR 
grants  from  the  NIH. 


266 


Reliability  of  Area  Mapping  by 
Delineation  in  Aerial  Photographs 

Claus-Peter1  Gross  and  Petra  Adler2 


Abstract.  -  In  a  test  series  under  standardised  conditions,  findings  between 
individual  photo-interpreters  concerning  forest  stand  boundaries 
delineation  were  investigated.  It  was  found  that  the  dispersion  of  the 
delineation  was  dependent  on  the  degree  of  difficulty  in  defining  a  natural 
border  by  the  interpreter.  The  evaluation  showed  that  90%  of  delineated 
polygon  lines  of  a  „normar  mapping  project  under  central  European 
forest  conditions  were  found  in  a  10  meter  buffer  with  a  photo  scale 
1:7000  and  required  22  meter  buffer  in  a  photo  scale  1:20000  to  fit  90% 
of  the  delineated  polygon  lines.  

INTRODUCTION 

Aerial  photographs  are  an  important  source  of  actual  and  detailed  area 
information.  For  decades  the  delineation  of  polygons  of  homogeneous 
information  from  images  for  forestry  or  landuse  application  has  been  common 
practice.  After  geometric  correction  using  some  kind  of  photogrammetric 
procedure  the  stand  polygons  can  be  integrated  as  an  information  layer  in  a 
Geographical  Information  System  (GIS). 

In  dependence  of  the  photogrammetric  solution  (low  budget,  advanced 
photogrammetry),  the  positional  error  of  the  polygon  lines  can  be  estimated. 
However,  there  is  still  uncertainty  concerning  the  scope  and  possible  variance  of 
interpretation,  because: 

•  the  photogrammetric  measurement  of  lines  is  less  precise  than  the 
measurement  of  point  objects. 

•  forests  are  natural  objects,  in  general  without  well  defined,  clear 
borderes. 

•  the  clearness  of  borders  between  different  forest  stands  varies  to  a  wide 
degree. 

Besides  the  photogrammetric  accuracy  the  interpreter  himself  is  responsible 
for  an  unknown  uncertainty.  It  is  of  great  interest  for  GIS  analysis  to  have  an 
quantitative  idea  of  the  influence  of  this  human  factor  on  the  spatial  data. 

1  Forester,  specialised  in  Remote  Sensing  (Ph.D.),  Dept.  of  Remote  Sensing  and  Landcsapinformationsystem, 
University  of  Freiburg,  Germany 

2  Student  of  Forestry  and  Remote  Sensing,  University  of  Freiburg.  Germany 


267 


MATERIAL  AND  METHODS 

In  the  study  aerial  Color-Infrared  (CIR)  photographs  of  two  different  scales 
were  evaluated.  The  large  scale  photographs  (1:7000)  show  an  intensively 
managed  forest  district  in  Germany  (Duvenhorst,  1995).  Small  scale  aerial 
photographs  (1 : 20000)  were  taken  from  a  forest  region  with  little  management 
activities  in  Italy  (Gross,  1993).  Both  photo-flights  were  taken  for  use  in  an 
intensive  forest  inventory  based  on  remote  sensing. 

In  a  first  step  the  degrees  of  difficulty  to  identify  the  forest  borders  was 
estimated  in  three  classes  (easy,  medium,  difficult)  using  a  representative 
check  of  all  polygon  lines  in  the  test  sites  deliniated  in  both  inventory  projects. 

In  a  second  step  the  reliability  of  the  forest  stand  delineation  was  tested. 
From  the  two  test  sites  a  representative  study  area  was  chosen  for  the  different 
photo  scale.  Ten  experienced  photointerpreters,  students  and  staff  members  of 
the  Department  had  to  delineate  five  forest  stand  boundaries  in  study  area 
using  defined  interpretation  keys.  For  the  mapping  procedure  a  high  precision 
photogrammetric  instrument  (ZEISS  Planicomp  P3)  was  used  to  exclude  the 
influence  of  equipment  errors. 

A  particular  problem  for  the  evaluation  of  the  test  series  is  the  lack  of 
reference  data.  The  borders  of  forest,  as  well  as  of  many  landscape  objects,  are 
„soft"  and  no  true  borderline  can  be  mapped,  using  either  terrestrial  means  or 
aerial  photographs.  To  handle  this  problem  adequately  and  to  quantify  the 
result  of  the  test,  a  special  GIS-  application  was  chosen  to  evaluate  the  test 
series: 

•  All  polygon  section  in  which  the  test  persons  easily  identified  the  same 
stand  border  were  cut  out  of  the  polygons  for  further  evaluations.  A 
fundamentaly  different  interpretation  of  a  forest  stand,  e.g.  the  division  of  a 
stand  in  smaller  unites,  was  not  of  central  interest  in  this  study.  These 
variations  can  be  influenced  through  training,  interpretation  guidelines  and 
the  experience  of  the  interpreters. 

•  Starting  from  the  minimal  and  maximal  polygon  line  of  the  entire  bundle, 
the  percentage  of  the  total  line  length  in  1  meter  (1:7000)  and  2  meter 
(1 : 20000)  buffers  was  calculated. 

•  The  raw  data  were  put  in  a  new  order,  starting  with  the  buffers  with  highest 
density  of  polygon  lines,  independent  of  their  localisation.  The  results 
presented  in  this  paper  show  the  average  between  the  results  for  maximum 
and  minimum  polygons  as  cumulative  histogram. 

RESULTS 

How  difficult  is  it  for  experienced  forest  photointerpreters  to  identify  and 
delineate  forest  stand  boundaries  from  aerial  photographs?  Independent  of  the 
scale  of  the  photographs  it  was  found,  that  about  2/3  of  all  stand  limits  were 


268 


Figure  1:  Degree  of  difficulty  to  identify  forest  stand  borders  in  aerial 
photographs  in  the  scale  of  1 :7000  (left)  and  1  20000  (right). 

clearly  visible  and  can  be  identified  without  any  doubt.  Only  about  20%  of  the 
forest  stands  were  found  to  be  extremely  difficult  to  delimit  to  neighbouring 
stands.  These  borders  were  found  to  be  more  or  less  fluid  (figure  1).  Of  course, 
the  visibility  might  change  with  natural  forest  conditions  and  silvicultural 
concepts.  For  the  situation  in  central  Europe  these  results  were  considered 
quite  representative. 

The  delineation  test  series  showed  the  following  results  for  the  two  photo 
scales: 

•  In  the  large  scale  photographs  the  variation  of  easy  to  identify  polygon 
lines  was  a  maximum  of  5  meters.  More  than  80%  of  all  lines  were  found  in 
a  buffer  of  2  meter  width.  The  delineation  of  medium  difficult  to  identify 
lines  varied  for  91%  in  a  11  meter  dispersion  whereas  a  buffer  width  of  27 
meter  was  necessary  to  fit  90%  of  all  lines  that  were  really  difficult  to 
interpret  (figure  2). 


c 

4) 

u 

I 

c 
0 
c 


1    2    3    4    5    6    7    8    9   10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  26  27 

buffer  width  in  meters 


Figure  2:  Variation  of  delineation  results  in  three  degrees  of  identification 
difficulty  (photo  scale  l  7000) 


269 


•  The  small  scale  photographs  naturally  show  a  higher  dispersion.  A  buffer 
of  10  meters  fits  90%  of  all  lines  that  were  easy  to  detect.  This  makes  0.5 
millimetres  (mm)  in  the  photo  scale  1:20000.  About  90%  of  the  medium 
difficult  stand  borders  were  found  in  a  38  meter  buffer  and  90%  of  the 
difficult  to  identify  lines  varied  in  a  46  meter  width  buffer.  To  fit  all  polygon 
lines  a  buffer  of  80  meter  was  necessary  (Figure  3).  The  precision  in 
coordinates  of  the  stereomodell  is  not  so  different  from  the  large  scale 
photos.  The  0.5mm  in  1.20000  for  easy  delineated  lines  is  similar  to  the 
0.4mm  in  1:7000.  The  difficult  to  interpret  lines  require  2.3mm  (1:20000) 
and  3  .8mm  (1:7000)  to  fit  90%  of  all  lines.  It  seems  that  the  small  scale 
photographs  may  have  even  some  advantages  for  the  interpretation  because 
of  the  better  overview. 


2    4    6    8   10  12  14  16  18  20  22  24  26  28  30  32  34  36  38  40  42  44  46  48 


buffer  width  in  meters 

Figure  3:  Variation  of  delineation  results  in  three  degrees  of  identification 
difficulty  (photo  scale  1:  20.000) 

CONCLUSION 

The  central  object  of  this  study  is  the  subjectivity  of  aerial  photo 
interpretation  related  to  spatial  accuracity  of  mapped  forest  stand  border  lines. 
Only  very  few  approaches  can  be  found  in  the  literature  to  quantify  the 
influence  of  the  human  factor  (e.g.  Congalton  1983,  Glemser  1993,  Scherrer 
et  al.  1990,  Zihlavnik  1991).  To  identify  the  correct  stand  border  well  worked 
and  defined  interpretation  keys  and  sufficient  training  of  the  interpreters  are 
preconditions.  But  even  when  the  interpreters  identify  the  same  stand  border 
and  are  willing  to  map  the  same  line  they  have  some  freedom  of  action.  The 
resulting  variations  are  strongly  related  to  the  clearness  of  the  stand 
delimitation.  The  photo  scale  has  mainly  an  effect  on  the  absolute  precision  in 
ground  coordinates. 


270 


When  the  delineation  variations  are  transferred  to  the  entire  inventory 
projects  by  weighting  the  data  with  the  portion  of  identification  difficulty,  then 
90%  of  all  polygon  lines  in  the  1:7000  photo  flight  were  in  a  buffer  width  of 
10  meters,  in  the  1:20000  photographs  a  width  of  22  meters  was  required. 
This  means,  speaking  from  the  point  view  of  a  forester,  under  normal 
conditions  the  accessible  precision  ranges  from  less  than  a  tree  crown  to 
maximal  one  adult  broadleaf  crown. 

Nevertheless  for  some  GIS  procedures  such  as  the  intersection  of  stand 
area  with  qualitative  data,  the  existing  uncertainy  should  be  taken  into 
consideration.  The  authors  propose  to  draw  buffers  with  the  same  probability 
around  the  mapped  polygon  lines  with  regard  to  the  difficulty  of  identification. 

REFERENCES 

Congalton,  R.  G.  1983:  A  Quantitative  Method  to  Test  for  Consistency  and 
Correctness  in  Photointerpretation.  Photogrammetric  Engineering  and 
Remote  Sensing  49  (1),  p.  69-74 

Duvenhorst,  J.  1995:  Photogrammetry  in  Forestry  -  Tool  for  Effective 
Inventory  and  Planning,  in  Fritsch,  D.;  Hobbie,  D.(eds  ):  Photogrammetric 
Week'95,  Wichmann  Heidelberg,  p.  337-349 

Glemser,  M.  1993:  Untersuchungen  zur  objektbezogenen  geometrischen 
Genauigkeit.  Salzburger  Geographische  Materialien,  Heft  20,  p.  97-108 

Gross,  CP.  1993:  Regionale  Waldinventur  zur  Erfassung  des  Waldzustandes 
mit  kleinmaCstabigen  Color-Infrarot  Luftbildern.  Dissertation  Universitat 
Freiburg,  126  p. 

Scherrer,  H.  V.;  Gautschi,  H.,  Hauenstein,  P.  1990:  Flachendeckende  Wald- 
zustandserfassung  mit  Infrarot-Luftbildern.  Bericht  der  Eidgenossischen 
Anstalt  fur  das  Forstliche  Versuchswesen  Nr.  318,  Birmensdorf,  101  p. 

Zihlavnik,  S.  1991:  Utilization  of  Photointerpretation  in  the  Determination  of 
Forest  Stand  Boundaries.  Lesnictvi,  p.  819-829 

BIOGRAPHICAL  SKETCH 

Claus-Peter  Gross  is  a  gratuated  forester  (Ph.D.)  After  five  years  practical 
work  in  the  forest  administration,  he  joined  in  1987  the  remote  sensing  team  at 
the  Forest  Faculty  in  Freiburg  (Prof.  Hildebrandt)  and  is  now  scientific 
employee  at  the  Department  of  Remote  Sensing  and  Landscape- 
Informationsystem  (Head  Prof.  B.  Koch).  He  is  specialised  in  forest  remote 
sensing  applications  and  photogrammetry. 

Petra  Adler  will  finish  her  studies  in  forestry  at  the  University  of  Freiburg 
1996.  In  her  studies  she  put  a  special  emphasis  on  remote  sensing. 


271 


The  Effect  Of  Database 
Generalization  On  The  Accuracy  Of 

The  Viewshed 

Peter  Fisher1 


Abstract  —  This  paper  examines  the  effects  of  database  generalization  on  the 
area  which  is  determined  to  be  visible  in  GIS  analysis.  Many  different  methods 
of  generalization  are  possible,  but  here,  for  any  cell  at  the  target  resolution, 
elevations  are  determined  from:  the  arithmetic  mean,  the  maximum,  the 
minimum,  and  the  maximum  difference  from  the  mean  of  the  cells  within  the 
kernel,  and  all  possible  combinations  of  regular  spacings.  The  resolutions 
studied  are  0.5,  0.33,  0.25,  and  0.2  of  the  original  study  area.  The  viewsheds 
determined  over  these  different  resolution  DEMs  are  compared  with  a  number  of 
possible  viewsheds  derived  by  generalization  of  the  viewshed  over  the  original 
DEM.  Of  those  tested  the  maximum  deviation  from  the  mean  within  kernel 
provides  the  best  estimate  of  the  pattern  and  area  of  the  viewshed  at  all 
resolutions. 


INTRODUCTION 

In  recent  years  considerable  attention  has  been  paid  to  two  related  lines  of 
research.  The  representation  of  digital  spatial  data  at  multiple  resolutions, 
where  either  information  is  stored  at  multiple  resolutions  or  is  generalized  from 
a  detailed  (large)  scale  to  a  generalized  (smaller)  scale  (Buttenfield  and 
McMaster  1991;  Muller,  Lagrange  and  Weibel  1995).  For  specific  studies  see, 
for  example,  Brown  et  al.  (1993),  Chang  and  Tsai  (1991),  Isaacson  and  Ripple 
(1990),  Joao  (1995)  and  Painho  (1995).  Few  pieces  of  work  have  taken  a 
dataset  at  one  single  resolution  and  examined  derived  products  in  alternative 
generalizations  of  the  data,  but  see  work  of  McMaster  (1987). 

The  research  reported  here  is  in  this  same  vein.  It  takes  the  viewshed  as  it  is 
determined  in  one  DEM  at  the  largest  resolution,  and  uses  measurements  of  the 
'quality'  or  'accuracy'  of  the  viewshed  when  it  is  determined  from  alternative 
derived  DEMs  at  smaller  resolutions.  The  purpose  is  to  comment  on  the 
changing  nature  of  the  viewshed  and  the  performance  of  the  alternative 
generalization  operators  defined. 


Senior  Lecturer,  University  of  Leicester,  Leicester,  United  Kingdom 


272 


THE  VIEWSHED 


The  viewshed  is  one  of  the  standard  procedures  included  with  GIS  designed 
for  analysis  of  elevation  data.  It  is  intended  to  distinguish  locations  which  can 
be  seen  from  a  particular  viewing  location  (in-view)  from  those  which  are  out- 
of-view.  A  line-of- sight  is  drawn  from  the  viewer  to  the  viewed  location,  and  if 
the  elevation  rises  above  the  elevation  of  the  line-of-sight  at  any  point  along  the 
line-of-sight  then  the  viewed  position  at  the  end  of  the  line  is  determined  to  be 
out-of-view.  Otherwise  it  is  in-view.  Here  the  method  of  point-to-point 
determination  across  a  rectangular  lattice  is  used  (Fisher  1993). 

EXPERIMENTAL  PROCEDURE 

For  the  research  reported  here  two  Ordnance  Survey  DEMs  from  the 
United  Kingdom  were  used:  one  of  the  Malvern  Hills  and  one  including  the 
southeast  part  of  Dartmoor.  Each  is  a  20  x  20  km  tile  of  the  national  elevation 
coverage  of  1 :50,000  gridded  elevation  data  recording  elevations  at  50  m 
intervals.  Therefore  each  is  401  x  401  values.  Within  each  test  area  100 
random  points  were  selected  within  an  active  area  which  excluded  a  60-cell 
buffer  from  the  edge  of  the  tile.  A  region  120  x  120  cells  was  then  taken 
around  each  point.  The  60,60  cell  within  each  area  was  taken  as  the  view  point 
and  the  viewshed  determined.  These  are  binary  products  which  are  referred  to 
below  as  the  original  viewsheds. 

Generalizations  were  then  made  of  the  DEM  itself,  the  original  viewshed 
and  the  viewpoint,  so  that  viewsheds  could  be  determined.  For  simplicity  the 
original  DEMs  are  generalized  so  that  pixels  in  the  derived  datasets  map  exactly 
to  the  original.  The  generalizations  included  0.5,  0.33,  0.25,  and  0.2  times 
reductions,  giving  DEM  arrays  of  60x60,  40x40,  30x30  and  20x20.  The 
generalization  procedures  are  discussed  in  the  next  section. 

Statistical  comparisons  include  the  bias  and  error  of  the  viewshed  in  the 
generalized  DEMs  was  derived  from  a  comparison  with  the  area  of  the  original 
viewshed,  and  the  kappa  coefficient  of  agreement  between  the  viewshed  in  the 
generalized  DEMs  and  a  number  of  alternative  generalizations  of  the  original 
viewshed. 

GENERALIZATION  OPERATORS 

Generalizing  the  DEM 

It  is  easy  to  envisage  many  alternative  approaches  to  the  generalization  of  a 
gridded  terrain  dataset.  The  methods  used  here  are  entirely  based  on  raster 
filtering.  They  do  not  include  any  surface  modelling  methods.  Another 
approach  is  to  resample  using  one  of  a  number  of  well  known  algorithms. 


273 


including  nearest  neighbour,  and  bilinear  interpolation.  These  methods, 
however,  are  generally  reserved  for  the  transformation  of  one  gridded  dataset 
to  another  where  the  orientation  of  the  grid  varies,  as  well  as  the  size  of  the 
grid.  Other  approaches  to  generalizing  the  grid  use  a  more  regular  geometric 
approach,  and  are  investigated  here. 

Two  primary  types  of  generalization  are  used.  The  first  type  searches  for 
values  in  the  original  DEM  on  a  regular  spacing  basis,  while  the  others  take 
summary  statistics  from  within  a  kernel  area. 

Regular  Spacing 

The  simplest  method  of  generalization  is  to  take  a  value  at  a  regular 
spacing,  w,  so  that  in  the  values  in  the  derived  DEM  are  every  m  th  values  in  x 
and  v  in  the  original  DEM  .  If  m  =  2  then  there  are  4  possible  realizations  of 
this  process.  zd  =  zh  then  zd  =  z2,  then  zd  =  z3,  and  finally zd  =  z4  (where  zu  z2, 
Z3,  and  z4  alternative  elevation  values,  and  zd  is  the  value  in  a  realization  of  the 
generalization  process).  If  m  =  5  then  there  are  25  different  possible  values  of 
zd. 

Summary  Statistics 

1)  Mean 

Within  the  kernel  area  the  most  obvious  summary  statistic  is  the  mean. 

2)  Maximum  and  Minimum 

Two  further  generalization  operators  are  the  maximum  and  minimum  values 
within  kernel. 

3)  Maximum  Deviation  from  the  Mean 

Finally,  the  value  of  the  elevation  which  is  the  largest  deviation  from  the 
mean  within  the  kernel  is  recorded. 

Generalizing  the  Viewshed 

For  comparison  with  derived  viewsheds  at  multiple  levels  of  generalization 
it  is  necessary  to  have  a  generalization  of  the  original  viewshed.  It  should  be 
recalled  that  the  viewshed  treated  here  is  a  binary  phenomenon,  any  cell  can 
only  be  treated  as  being  in- view  or  out-of-view.  Three  different  strategies  are 
used: 

1)  All 

A  cell  in  the  generalized  dataset  is  taken  to  be  in-view  if  all  cells  in  the 
kernel  are  in-view.  Otherwise  it  is  out-of-view. 

2)  Majority 

A  cell  in  the  generalized  dataset  is  taken  to  be  in-view  if  the  majority  of  cells 
in  the  kernel  are  in-view.  Otherwise  it  is  out-of-view.  Where  there  is  an  even 
number  of  cells  in  the  kernel  the  majority  is  taken  as  n/2  +1 . 

3)  Any 


274 


A  cell  in  the  generalized  dataset  is  taken  to  be  in-view  if  the  any  cell  in  the 
kernel  is  in-view.  Otherwise  the  generalized  cell  is  out-of-view. 

Generalizing  the  Viewpoint 

Finally,  the  treatment  of  the  viewpoint  is  considered.  Viewshed  algorithms 
treat  the  viewpoint  very  differently.  Here,  the  cell  considered  to  be  the 
viewpoint  in  any  generalization  is  the  generalized  cell  which  contains  the 
viewpoint  in  the  ungeneralized  DEM.  This  approach  has  the  advantage  over 
others  that  the  viewpoint  is  always  at  the  centre  of  a  cell  in  any  dataset,  and  so 
coincides  geometrically  with  a  height  value  in  the  corresponding  DEM.  The 
DEM  elevation  at  the  viewpoint  is  always  known  as  part  of  the  generalized 
DEM  and  never  inferred.  Many  possible  reslizations  of  the  original  viewshed 
correspond  to  the  viewshed  at  a  particular  scale,  and  so  results  here  must  be 
regarded  as  provisional. 

Implementation 

The  above  approach  was  implemented,  together  with  the  viewshed 
algorithm  in  Turbo  Pascal  7.0.  The  sampling,  generalization  by  all  methods, 
generation  of  the  viewshed  and  calculation  of  summary  and  comparitive 
statistics  of  the  viewsheds  were  all  integrated.  It  ran  on  a  Pentium  PC  which 
had  been  checked  for  faults  on  the  processor.  Idrisi  v  4. 1  was  used  in  support 
of  the  analysis  to  view  data  at  different  resolutions,  etc.  and  analysis  was  done 
in  Excel  5.0  and  using  bespoke  programs. 

RESULTS 

Generalizing  the  Original  Viewshed 

As  noted  above,  three  different  versions  of  generalizing  the  original 
viewshed  were  derived.  These  are  based  on  all.  a  majority  and  any  of  the 
pixels  within  the  kernel  being  in  view  in  the  original.  The  results  of  this  analysis 
can  best  be  seen  in  Table  1 ,  where  the  correlation  between  the  area  of  the 
original  and  generalized  viewsheds  is  reported  together  with  the  mean  bias  and 
error  as  a  result  of  the  generalization  for  both  test  areas.  It  can  be  seen  that  the 
correlation  for  the  majority  operator  is  always  closes  to  1 .0,  and  that  it  has 
much  the  lowest  values  of  bias  and  error  at  all  levels  of  generalization.  In  short, 
the  area  of  the  viewshed  derived  by  the  majority  operator  is  predictable,  and  is 
very  similar  to  the  original  area,  even  over  large  resolution  changes,  while  the 
values  as  a  result  of  the  any  and  all  operators  are  not  so  predictable  at  the  larger 
resolution  changes.  Furthermore,  the  values  of  the  majority  operator  are  very 
close  to  the  values  for  the  original. 


275 


Table  1.  —  Bias  and  Error  in  generalizations  of  the  Orignal  Viewshed. 


Generalizations 

IVTsiIvprn 

Diirtmnnr 

Bias 

Error 

Correl. 

Bias 

Error 

Correl. 

0.5  x 

Any 

3.39 

3  78 

0  993 

\J  .  y  y  ~J 

2  42 

2  70 

0  997 

Major 

-1  19 

1  •  1  jr 

1  34 

0.997 

-0  87 

0  97 

\J  *y  1 

0  999 

\J  ,y  y  y 

All 

-3.60 

4.06 

0.985 

-2.48 

2.79 

0.993 

0.33  x 

Any 

■J 

5.98 

6  58 

0  983 

4  50 

4  97 

0  993 

\j .  y  y  ~j 

0  09 

0  43 

0  999 

\j .  y  y  y 

0  00 

0  24 

0  999 

\J  *y  y  y 

All 

£  0  1 
-O.Z  1 

/.uz 

-4.43 

J.U4 

C\  07") 

0.25  x 

Any 

8.36 

9.12 

0.971 

6.46 

7.10 

0.988 

Major 

-0.56 

0.77 

0.997 

-0.50 

0.63 

0.998 

All 

-7.93 

8.99 

0.881 

-5.96 

6.75 

0.940 

0.2  x 

Any 

10.54 

11.46 

0.958 

8.30 

9.09 

0.980 

Major 

-0.29 

0.70 

0.997 

-0.26 

0.61 

0.995 

All 

-9.13 

10.39 

0.803 

-6.99 

7.92 

0.896 

The  Area  of  the  Viewshed 

When  the  viewshed  is  determined  over  the  generalized  DEMs,  there  are  a 
number  of  ways  to  compare  it  with  the  original  viewshed.  Two  parameters  are 
examined  here:  the  area  visible,  and  the  kappa  coefficient  of  agreement.  The 
area  of  the  viewshed  at  different  resolutions  is  easily  determined,  and  in  Table  2 
the  mean  bias  and  error  are  given  for  both  test  areas  at  all  four  resolutions.  In 
addition,  the  correlation  coefficients  between  the  two  areas  are  reported.  The 
"best"  generalization  is  that  which  yields  the  smallest  bias  and  error  and  the 
largest  correlation. 

Table  2  is  divided  by  the  amount  of  and  methods  of  generalization.  The 
values  reported  for  regular  spacings  are  the  riiinimum  area  and  the  maximum 
area  determined  among  the  4,  9,  16  and  25  possible.  There  is  considerable 
variation  among  these,  as  is  reflected  in  the  error  measures  which  are  always 
the  largest  and  smallest  of  any  other  generalization.  On  the  other  hand,  the 
correlation  coefficients  for  these  are  on  the  whole  both  poorer  than  the  best  of 
the  summary  statistic  values. 

Among  the  summary  statistics  of  the  generalizations,  21  out  of  the  24 
measures  of  accuracy  show  the  maximum  deviation  from  the  mean  to  give  the 
best  performance.  The  correlation  coefficients  and  mean  errors  yielded  by  this 
method  are  always  the  best  of  those  reported.  The  bias  is  sometimes  smaller 
for  some  other  method  (in  two  cases  the  maximum  within  kernel  and  once  the 
minimum  within  kernel). 


276 


Table  2.  —  Error  measures  for  the  area  of  the  viewshed  in  generalizations  of 
the  vjewshed  as  compared  with  the  area  in  the  original  viewshed.  


Generalizations 

Malvern 

Bias        Error  Correl 

Dartmoor 

Bias        Error  Correl. 

0.5  x      regular  Max 
spacing  Min 

6.17      8.17  0.802 
1.14      3  32  0.913 

3.60      4.34  0.956 
0.47       1.82  0.965 

summary  Max 
statistics  Min 
Mean 
MaxDev 

4.04      6.56  0.809 
4.27      6.63  0.823 
4.52      6.76  0.830 
3.39      4.02  0.967 

2.18      3.18  0.946 
2.03      3.12  0.957 
2.23      2.97  0.972 
2.11       2.58  0.992 

0.33  x   regular  Max 
spacing  Min 

11.39     13.68  0.706 
2.28      6.42  0.725 

7.52      8.75  0.871 
1.03      3.33  0.897 

summary  Max 
statistics  Min 
Mean 
MaxDev 

6.85      9.34  0.768 
8.51     11.25  0.688 
7.87     10.10  0.768 
6.60      8.08  0.884 

4.50      6.05  0.874 
4.57      6.59  0.880 
4.74      6.17  0.910 
4.08      4.85  0.979 

0.25  x    regular  Max 
spacing  Min 

16.34     18.89  0.600 
3.34      7.42  0.673 

12.26     14.26  0.712 
1.76      4.88  0.812 

summary  Max 
statistics  Mm 
Mean 
MaxDev 

9.55     12.24  0.658 
12.61     15.86  0.600 
11.22     14.01  0.675 

9.65     11.23  0.844 

7.46      9.40  0.784 
7.25      9.48  0.806 
7.38      9.12  0.836 
6.29      7.53  0.955 

0.2  x     regular  Max 
spacing  Min 

20.75     23.23  0.526 
4.15      8.08  0.651 

16.60     18.87  0.648 
2.30      5.51  0.738 

summary  Max 
statistics  Min 
Mean 
MaxDev 

12.45     15.88  0.542 
15.83     18.96  0.580 
14.00     17.12  0.607 
12.77     14.60  0.790 

9.78     12.20  0.705 
10.26     12.62  0.685 
10.14     12.38  0.706 

8.97     10.57  0.927 

The  latter  of  these  shows  an  enormous  spread  of  values  which  tends  to  be 
more  typical  of  the  0.2  generalizations  than  that  yielded  by  the  maximum 
deviation  from  the  mean.  The  0.5  generalizations  are  generally  relatively  well 
correlated  as  is  shown  by  the  correlation  coefficients,  but  this  can  rapidly 
deteriorate. 

The  Arrangement  of  the  Viewshed 

The  kappa  coefficient  of  agreement  between  two  sets  of  data  ranges  from  0 
to  1 .  where  1  reflects  perfect  correspondence  in  the  arrangements  and  0 
disagreement.  It  has  been  widely  applied  to  examining  the  confusion  matrix  in 
remote  sensing  (Congalton  et  al.,  1983).  Here  the  same  approach  is  used. 
Kappa  is  determined  for  2  x  2  tabulations  which,  for  a  particular  resolution, 
compares  the  visbility  of  pixels  in  the  generalized  binary  viewshed.  as 
determined  by  the  majority  method,  with  that  determined  through  the  DEM 
when  generalized  by  one  of  the  methods  under  test.  Summary  results  are 


277 


reported  in  Table  3  where  the  maximum,  minimum  and  mean  values  of  the 
coefficient  for  the  1 00  viewpoints  in  each  study  area  are  reported.  The 
maximum  and  minimum  values  are  reporting  the  best  and  the  worst  agreements 
between  patterns  in  any  situation.  The  best  generalization  method  should  yield 
the  largest  value  of  kappa  at  any  reduction. 

The  results  for  regularly  spaced  samples  report  the  agreements  for  the 
maximum  and  minimum  areas  in  any  single  generalization  test.  The  maximum 
viewshed  is  the  best  in  22  out  of  the  24  values  reported.  In  other  words,  the 
largest  viewshed  generated  by  regular  spacing  from  the  kernel  area  yields  the 
best  agreement  in  the  spatial  pattern  of  the  viewshed  among  all  other 
generalizations  of  the  DEM.  On  the  other  hand,  the  minimum  area  viewshed 
always  yields  the  worst  agreement.  This  method  of  generalization  provides  an 
envelope  of  all  other  values.  Indeed,  in  the  0.2  generalizations  of  both  the 
Malvern  and  Dartmoor  test  areas,  regular  spacing  yielding  the  minimum  visible 
area  gives  almost  total  disagreement  with  the  original  viewshed  (0.054  and 
0.020). 

Table  3.  —  Kappa  coefficients  of  agreement  between  the  generalized  version 


of  the  original  viewshed  and  the  viewshed  determined  over  the  generalized  DEM. 


VXC11CI  luu<<lllUI19 

lVfiilvprn 

Max        Mean  Min 

Durtmnnr 

Max       Mean  Min 

0.5  x     regular  Max 
spacing  Min 

0.899  0.744  0.285 
0.850     0.591  0.201 

0.924  0.818  0.661 
0.845     0.688  0.135 

summary  Max 
statistics  Min 
Mean 
MaxDev 

0.878  0.674  0.231 
0.910  0.661  0.217 
0.902  0.692  0.222 
0.899     0.725  0.263 

0.903  0.760  0.371 
0.879  0.769  0.414 
0.890  0.784  0.411 
0.924    0.802  0.440 

0.33  x    regular  Max 
spacing  Min 

0.886  0.694  0.235 
0.805     0.465  0.106 

0.877  0.772  0.480 
0.796    0.540  0.162 

summary  Max 
statistics  Min 
Mean 
MaxDev 

0.868  0.607  0.182 
0.856  0.562  0.158 
0.873  0.624  0.193 
0.864     0.646  0.196 

0.864  0.666  0.201 
0.853  0.675  0.316 
0.858  0.705  0.278 
0.877     0.728  0.480 

0.25  x    regular  Max 
spacing  Min 

0.856  0.601  0.170 
0.748     0.353  0.053 

0.864  0.688  0.427 
0.695     0.406  0.053 

summary  Max 
statistics  Min 
Mean 
MaxDev 

0.833  0.496  0.106 
0.816  0.447  0.102 
0.831  0.510  0.133 
0.812     0.528  0.118 

0.779  0.559  0.126 
0.795  0.555  0.245 
0.829  0.593  0.180 
0.795    0.615  0.379 

0.2  x     regular  Max 
spacing  Min 

0.833  0.555  0.174 
0.711     0.288  0.054 

0.827  0.634  0.238 
0.613     0.320  0.020 

summary  Max 
statistics  Min 
Mean 
MaxDev 

0.785  0.446  0.082 
0.785  0.393  0.015 
0.825  0.458  0.109 
0.750     0.469  0.118 

0.813  0.487  0.047 
0.775  0.484  0.067 
0.815  0.528  0.059 
0.767    0.533  0.193 

278 


In  1 7  out  of  24  summaries,  the  maximum  deviation  within  kernel  yields  the 
best  agreement.  In  5  cases  the  mean  is  best,  and  the  maximum  and  minimum 
within  kernel  are  best  once  each.  With  0.2  generalization  the  worst  levels  of 
agreement  are  again  very  poor  (0.01 5  in  the  0.2  generalization  of  the  Malvern 
area  by  minimum  within  kernel).  Both  the  mean  and  the  best  levels  of 
agreement  in  all  situations  are,  however,  quite  acceptable. 

CONCLUSION 

A  number  of  conclusions  can  be  drawn  from  the  current  work.  For  the 
viewshed  it  would  appear  that  the  generalization  of  the  binary  viewshed  to 
other  grid  resolutions  is  best  performed  by  determining  those  cells  at  the  target 
resolution  with  the  majority  of  visible  cells.  Other  immediately  apparent 
methods  of  generalization  do  not  yield  results  which  match  the  area  and  pattern 
of  the  original  viewshed. 

Generalization  of  the  DEM  yields  many  alternative  possible  viewsheds. 
Regular  spacing  of  the  kernel  area  yields  both  the  best  and  worst  estimates  of 
the  visible  area.  The  result  is  therefore  unpredictable,  and  although  the  method 
is  fast  and  convenient  it  is  not  to  be  recommended  as  a  basis  of  generalization. 
Among  the  statistical  summaries  of  a  kernel  area,  generalization  of  the  DEM  by 
determination  of  the  maximum  deviation  within  the  kernel  yields  the  viewshed 
which  best  reflects  both  the  area  and  arrangement  of  the  visible  area.  The 
results  are  all  most  stable  and  most  frequently  the  best  for  this  method.  It  is 
therefore  the  method  to  be  recommended  (of  those  tested)  in  any  situation 
where  a  DEM  requires  generalization,  and  the  viewshed  is  the  derived  product 
which  is  uppermost  in  the  investigator's  interest. 

Finally  it  should  be  noted  that  in  most  cases  neither  the  area  not  the  pattern 
of  the  viewshed  is  badly  disrupted  by  generalization  of  the  DEM,  although  it 
can  be  if  an  injudicious  method  of  generalization  is  used. 

As  stated  in  the  introduction,  very  little  research  is  reported  on  propagating 
the  effects  of  alternative  generalizations  of  spatial  databases.  Research  similar 
to  that  reported  here  can  be  envisaged  for  almost  any  spatial  data  and  derived 
product.  Indeed,  it  is  crucial  for  very  many  applications  that  we  know  best  how 
to  generalize  both  categorical  data  (the  binary  viewshed),  and  continuous  data 
(the  DEM),  and  possible  consequences  of  generalization.  While  the  results 
reported  here  seem  quite  conclusive,  and  are  very  likely  to  be  generalizable  to 
other  areas,  so  long  as  the  viewshed  is  of  interest,  there  is  absolutely  no 
guarantee  that  either  the  recommended  method  of  generalization  should  be  the 
same  for  all  derived  products,  or  that  there  is  not  a  better  generalization 
operator  for  the  viewshed. 


279 


ACKNOWLEDGEMENTS 


I  would  particularly  like  to  thank  Jo  Wood  for  some  insightful  suggestions 
in  the  course  of  this  work. 

REFERENCES 

Brown,  D.G.,  Ling  Bain  and  Walsh,  S.J.,  1993.  Response  of  a  distributed 

watershed  erosion  model  to  variations  in  the  input  data  aggregation  levels. 

Computers  &  Geosciences  19,  499-509. 
Buttenfleld,  B.P.,  andMcMaster,  R.B.  (Editors)  1991.  Map  Generalization: 

Making  Rules  for  Knowledge  Representation.  London:  Longman. 
Chang,  K.-t.,  and  Tsai,  B.-w.  1991.  The  effect  of  DEM  resolution  on  slope  and 

apect  mapping.  Cartography  and  Geographic  Information  Systems  18,  69- 

77. 

Congalton,  R.G.,  Oderwald,  R.G.,  and  Mead,  R.A.,  1983.  Assessing  Landsat 

classification  accuracy  using  discrete  multivariate  analysis  statistical  techniques. 

Photo grammetric  Engineering  and  Remote  Sensing  49,  79-87. 
Fisher,  P.F.  1 993 .  Algorithm  and  Implementation  Uncertainty  in  Viewshed 

Analysis.  International  Journal  of  Geographical  Information  Systems  7,  331- 

347. 

Isaacson  ,  D.L.,  and  Ripple,  W.J.  1990.  Comparison  of  7.5-minute  and  1- 
degree  digital  elevation  models.  Photo  grammetric  Engineering  and  Remote 
Sensing  56,  1523-1527. 

Joao,  E.M.  1995.  The  importance  of  quantifying  the  effects  of  generalization. 
In  GIS  and  Generalization:  Methodology  and  Practice,  edited  by  Muller, 
J.C.,  Lagrange,  J.P.  and  Weibel,  R.  London:  Taylor  &  Francis.pp.  183-193. 

McMaster,  R.B.  1987.  The  geometric  properties  of  numerical  simplification. 
Geographical  Analysis  19,  330-346. 

Muller,  J.C.,  Lagrange,  J.P.  and  Weibel,  R.  (Editors)  1995.  GIS  and 

Generalization:  Methodology  and  Practice.  London:  Taylor  &  Francis.  257 

P- 

Painho,  M.  1995.  The  effects  of  generalization  on  attribute  accuracy  in  natural 
resource  maps.  .  In  GIS  and  Generalization:  Methodology  and  Practice, 
edited  by  Muller,  J.C.,  Lagrange,  J.P.  and  Weibel,  R.  London:  Taylor  & 
Francis.pp.  194-206. 


280 


Modeling  Slope  Stability  Uncertainty:  A 
Case  Study  at  the  H.J.  Andrews  Long-Term 
Ecological  Research  Site,  Oregon 

Michelle  L.  Murillo1  and  Gary  J.  Hunter2 


Abstract. — A  slope  stability  model  employing  Digital  Elevation 
Model  (DEM)  data  was  used  in  a  Geographic  Information  System 
(GIS)  to  predict  landslide  susceptibility  in  a  large  region  of  the  H.J. 
Andrews  Long-Term  Ecological  Research  (LTER)  site,  located  in  the 
Western  Cascades  of  Oregon.  To  assess  the  uncertainty  of  the  final 
output,  several  different,  but  equally  probable,  versions  of  the  input 
DEM  were  created  through  the  addition  of  random,  spatially 
autocorrelated  noise  (error)  files.  The  DEMs  were  then  processed  to 
produce  a  family  of  slope  stability  maps  from  which  the  uncertainty 
effects  of  DEM  elevation  error  upon  the  final  susceptibility  index 
could  be  assessed.  The  slope  stability  model  itself  involved  the 
overlay  of  rock  strengths  with  slope  classes  to  establish  landslide 
susceptibility  indices.  The  inherent  nature  of  error  in  DEMs  used  for 
modeling  slope  stability,  coupled  with  much  of  the  Pacific  Northwest's 
commercial  timber  being  located  in  mountainous  terrain,  makes  this 
an  intriguing  problem  to  solve.  The  ability  to  understand  the  resultant 
uncertainty  due  to  elevation  error  when  applied  to  the  model  has  the 
potential  to  facilitate  improved  natural  resource  management 
decisions  in  relation  to  harvesting  and  subsequent  slope  stability. 


INTRODUCTION 

The  US  Pacific  Northwest  contains  some  of  the  most  productive  temperate 
forests  in  the  world  and  timber  harvesting  has  been  a  primary  land  use  in  the  area 
for  over  100  years.  Because  most  of  the  commercial  timber  is  in  mountainous 
terrain,  such  land  use  activities  tend  to  decrease  slope  stability  and  therefore 
increase  the  rate  of  mass  movements.  One  of  the  major  concerns  associated  with  an 
increase  in  such  movement  is  the  cumulative  effects  it  has  on  the  entire  watershed — 
such  as  decreased  water  quality,  loss  of  spawning  habitat  and  organic  matter,  and 
debris  jams  that  may  break  during  peak  flows,  thereby  scouring  channels  and 
destroying  riparian  vegetation.  Thus,  the  ability  to  predict  slope  failure  is  a  valuable 
resource  in  determining  the  impacts  of  forestry  practices  (Swanson  et.  al.  1981). 

While  methods  such  as  ground  and  aerial  surveys  can  aid  in  locating  existing 
and  potential  landslide  areas,  they  also  suffer  several  disadvantages.  For  example, 
in  the  Western  Cascade  range,  ground  surveys  are  difficult  to  carry  out  since 
accessibility  is  a  major  problem  in  a  region  characterized  by  steep  slopes  and  very 
few  roads  (Swanson  et  al.  1981).  This  effectively  reduces  the  amount  of  terrain 
that  can  be  covered,  which  in  turn  may  lead  to  an  underestimation  of  landslide 
occurrences. 


Graduate  student,  Department  of  Forest  Science,  Oregon  State  University.  Conallis.  OR 
Assistant  Professor,  Department  of  Geomatics.  The  University  of  Melbourne.  Victoria.  Australia 


281 


Similarly,  there  can  be  difficulties  in  obtaining  suitable  aerial  photography  as  a 
result  of  highly  variable  weather  and  climate  conditions,  seasonal  considerations, 
and  problems  with  photographic  interpretation.  Unfortunately,  landslide 
inventories  in  the  Western  Cascades  are  best  conducted  during  the  winter  rainy 
season  and/or  immediately  after  major  storms  when  cloud  cover  is  most  prevalent. 
Also,  landslides  may  be  too  small  to  be  detected  by  aerial  photography  (Swanson 
and  Lienkaemper  1985);  or  difficult  to  identify  because  they  are  either  hidden  under 
the  forest  canopy  (McKean  et.  al.  1991),  masked  in  shadow  by  steep,  narrow 
ridges  and  tall,  dense  forest  vegetation  (Skaugset  1992);  or  located  on  streamsides. 

Alternatively,  we  can  use  Geographic  Information  Systems  (GIS)  to  help  model 
landslide  susceptibility  by  applying  the  critical  factors  used  in  determining  slope 
stability.  While  this  approach  permits  large-scale  evaluation  of  the  landscape,  it 
suffers  from  a  lack  of  knowledge  about  the  uncertainty  of  the  final  output  resulting 
from  errors  occurring  in  the  parameters  used  to  model  potential  slope  failure.  The 
purpose  of  this  paper  is  to  describe  how  the  application  of  a  slope  stability  model 
based  on  DEM-<lerived  values  was  enhanced  to  assess  the  uncertainty  in  predicting 
slope  failure  in  the  H.J.  Andrews  Long-Term  Ecological  Research  (LTER)  site. 
The  model  overlays  slope  gradient  values  with  a  generalized  geology  map  to  yield 
landslide  susceptibility  classes. 

In  order  to  apply  the  procedure  to  the  slope  stability  model,  the  uncertainty 
assessment  technique  described  in  Hunter  and  Goodchild  (1995)  was  employed  in 
which  different,  but  equally  probable,  versions  of  the  source  DEM  were  created 
through  the  addition  of  random,  spatially  autocorrelated  elevation  error  files.  The 
DEMs  were  then  processed  to  produce  a  family  of  slope  stability  maps  from  which 
the  effects  of  error  in  DEM  elevations  upon  the  final  landslide  susceptibility  indices 
could  be  assessed.  This  paper  is  structured  such  that  a  description  of  the 
uncertainty  model  is  presented  next,  followed  by  discussion  of  the  application  of 
the  method  to  the  slope  stability  model,  and  finishing  with  some  closing  remarks. 

DISCUSSION  OF  THE  UNCERTAINTY  MODEL 

Before  discussing  the  uncertainty  model  used  in  this  research,  some  explanatory 
remarks  are  required  regarding  the  term  'uncertainty'  itself.  In  general  terms,  it 
denotes  a  lack  of  sureness  or  definite  knowledge  about  an  outcome  or  result  and,  in 
the  context  of  GIS,  we  suggest  there  is  a  clear  distinction  between  'error'  and 
'uncertainty',  since  the  former  implies  that  some  degree  of  knowledge  has  been 
attained  about  differences  between  actual  results  or  observations  and  the  truth  to 
which  they  pertain.  On  the  other  hand,  'uncertainty'  conveys  the  fact  that  it  is  the 
lack  of  such  knowledge  which  is  responsible  for  our  hesitancy  in  accepting  those 
same  results  or  observations  without  caution,  and  the  term  'error'  is  often  used 
when  it  would  be  more  appropriate  to  use  'uncertainty'. 

The  uncertainty  model  that  has  been  used  here  is  a  version  of  the  one  originally 
developed  by  Goodchild  et  al.  (1992).  In  general  terms  it  may  be  defined  as  a 
stochastic  process  capable  of  generating  a  population  of  distorted  versions  of  the 
same  reality,  with  each  version  being  a  sample  from  the  same  population.  The 
traditional  Gaussian  model  (where  the  mean  of  the  population  is  an  estimate  of  the 
true  value  and  the  standard  deviation  is  a  measure  of  the  variation  in  observations) 
is  one  attempt  at  describing  error,  but  it  says  nothing  about  local  variation  or  the 


282 


processes  by  which  it  has  accumulated.  The  model  applied  here  is  viewed  as  an 
advance  on  the  simple  Gaussian  approach  since  it  has  the  ability  to  show  spatial 
variation  in  uncertainty,  as  well  as  the  capacity  to  include  in  its  realizations  the 
probable  effects  of  error  propagation  resulting  from  the  various  algorithms  and 
processes  that  have  been  progressively  applied  to  the  data  sets  employed.  By 
studying  different  versions  of  the  final  output,  it  is  possible  to  see  how  differences 
in  input  affect  the  outcome,  and  in  essence  the  purpose  of  the  model  could  be 
described  as  an  attempt  to  "find  a  Gaussian  distribution  for  maps". 

Elsewhere,  in  a  paper  by  Hunter  and  Goodchild  (in  review),  it  has  been  argued 
that  while  it  is  possible  to  distort  or  perturb  a  data  set  according  to  an  error 
description  (such  as  the  Root  Mean  Square  Error  (RMSE)  value  for  a  DEM) 
without  any  consideration  of  the  likely  spatial  autocorrelation  between  point  sample 
elevations,  the  process  may  be  stochastic  but  inevitably  lacks  a  certain 
'truthfulness' — since  adjacent  elevations  in  a  DEM  which  are  otherwise  similar  in 
value  can  be  severely  distorted,  thereby  creating  large  pits  and  peaks  which  often  do 
not  intuitively  occur  in  nature.  This  approach  produces  what  are  known  as  'random 
maps'. 

On  the  other  hand,  the  assumption  of  complete  spatial  dependence  between 
neighbouring  points  produces  realizations  of  a  DEM  which  appear  'truthful'  but  not 
stochastic,  since  elevations  are  unnaturally  constrained  to  maintain  their  relative 
differences  to  each  other  and  the  introduction  of  a  noise  (or  error)  component  has 
the  effect  of  moving  all  DEM  elevations  up  or  down  by  a  constant  amount.  Hence, 
there  is  a  need  to  find  the  value  in  the  domain  0  <  p  <  0.25  (where  p  is  a  measure  of 
spatial  autocorrelation)  which  meets  the  requirements  of  being  both  stochastic  and 
'truthful'.  The  limit  of  0.25  ensures  stationarity  (as  discussed  in  Cliff  and  Ord, 
1981,  p.  147)  when  the  Rook's  case  is  used  to  test  a  cell's  elevation  against  its  four 
neighbours  sharing  a  common  edge. 

Application  of  the  uncertainty  model  consists  of  four  stages  (as  described  more 
fully  in  Hunter  and  Goodchild,  1995),  with  the  first  stage  requiring  the  user  to 
combine  whatever  data,  processes  and  models  are  needed  to  generate  the  desired 
output — in  other  words,  applying  the  GIS  as  would  normally  occur  without  any 
consideration  of  uncertainty.  Secondly,  the  parameters  necessary  for  the  realization 
process  are  determined  by  reading  system  variables  associated  with  the  source 
DEM,  such  as  the  number  of  rows  and  columns  in  the  data  file,  the  cell  size,  and 
geo-referencing  details.  These  will  be  required  later  when  the  elevation  noise  files 
are  transformed  to  agree  with  the  original  DEM.  An  error  estimate  for  the  DEM 
will  also  need  to  be  identified  and  this  is  usually  a  global  value  of  the  elevation  error 
present,  such  as  the  RMSE  as  supplied  by  the  DEM  producer. 

Whilst  not  a  direct  step  in  the  realization  procedure,  the  noise  files  to  be 
employed  are  usually  pre-computed  and  permanently  stored  in  the  system  for  future 
use.  To  date,  it  has  been  considered  sufficient  in  the  applications  tested  for  about 
ten  files  to  be  held  against  each  p  value,  with  the  default  values  being  p  =  0.0,  0.05, 
0.10,  0.15,  0.20,  0.21,  0.22,  0.23,  0.24,  0.245,  and  0.249.  As  for  the  maximum 
value  of  p  employed  (0.249),  experience  has  shown  there  is  little  to  be  gained  from 
using  values  higher  than  this  since  the  realization  process  becomes  so  constrained 
that  there  is  no  discernible  difference  between  the  realized  maps  and  the  original 
product. 


283 


In  stage  3  of  the  methodology,  it  is  expected  that  users  will  want  to  see  a  small 
number  of  trial  realizations  and  the  default  values  of  p  listed  above  are  applied.  A 
single  realization  for  each  value  is  obtained  by  first  applying  the  parameters  derived 
from  stage  2  to  geo-reference  and  transform  the  coordinates  of  the  noise  grid. 
Next,  the  error  estimate  is  applied  to  map  the  noise  values  from  a  normal 
distribution  of  N(0,1)  to  N(0,RMSE).  This  adjusted  noise  file  is  added  to  the 
source  data  to  produce  a  realization  to  which  the  commands  employed  to  create  the 
original  GIS  product  are  applied.  The  realized  maps  and  the  differences  between 
the  realizations  and  the  original  output  can  be  displayed  in  map  or  graph  form. 
Finally,  in  stage  4  of  the  process  the  user  focuses  on  a  specific  p  value  in  the  range 
0  <  p  <  0.25,  and  applies  a  series  of  noise  realizations  to  produce  a  set  of  distorted 
DEMs  for  which  a  family  of  outputs  are  created.  These  enable  assessment  of  the 
uncertainty  in  the  final  product  due  to  elevation  error  in  the  original  DEM. 

THE  CASE  STUDY  AND  DISCUSSION  OF  RESULTS 

The  H.J.  Andrews  Experimental  Forest  is  located  approximately  80  km  east  of 
Eugene,  Oregon.  Lookout  Creek  Basin  (the  test  area)  lies  within  the  forest  and 
drains  nearly  6400  hectares  in  the  Western  Cascade  Range.  The  overall  site  was 
established  as  an  Experimental  Forest  in  1948  after  which  trial  harvests  took  place 
and  concomitant  road  construction  occurred  throughout  the  1960s.  Since  1970, 
site  emphasis  has  shifted  to  ecosystem  research  and  the  levels  of  timber  harvesting 
and  road  building  have  been  reduced.  The  upper  areas  of  Lookout  Creek  Basin  are 
formed  from  lava  flows  ranging  in  age  from  3  to  13  million  years,  while  lower 
areas  are  underlain  by  older  (14  to  25  million  years  BP),  more  weathered  and 
hydrothermally  altered  clastic  volcanic  rocks  (Swanson  et.  al.  unpublished). 

Two  types  of  landslides  occur  in  the  basin.  The  most  common  are  relatively 
small  (averaging  approx.  2000  m3),  shallow,  rapid  debris  slides  that  take  place  on 
hillslopes  during  intense  rainfall  and  rain-on-snow  events.  The  second  type  are 
relatively  large,  deep-seated,  slow-moving  landslides  (earthflows)  that  move 
seasonally  at  rates  varying  from  centimeters  to  meters  per  year.  An  extensive 
inventory  was  conducted  which  documented  over  140  debris  slides  since  1950  in 
the  basin  (Dyrness  1967,  Swanson  and  Dyrness  1975,  Swanson  et.  al.  1981). 
Landslide  susceptibility  in  the  basin  has  been  interpreted  to  be  a  function  of  both 
slope  steepness  and  soil  strength,  as  represented  by  the  type  of  underlying  bedrock 
(Dyrness  1967,  Swanson  and  Dyrness  1975). 


Table  1. — Derivation  of  landslide 
susceptibility  classes. 

Slope      Rock  Strength  Susceptibility 


0°  <  10° 

Weak 

Moderate 

Moderate 

Low 

Strong 

Low 

10° -20° 

Weak 

High 

Moderate 

Moderate 

Strong 

Low 

>20° 

Weak 

High 

Moderate 

High 

Strong 

Low 

In  applying  GIS  to  produce  a 
landslide  susceptability  index  map,  a 
DEM  produced  by  the  U.S.  Geolog- 
ical Survey  (USGS)  was  used  which 
has  a  30-m  resolution  and  an  elevation 
RMSE  of  7  meters.  The  basin  is 
located  within  a  portion  of  the  DEM 
measuring  309  x  426  cells  and 
occupies  some  69,952  cells  or  53%  of 
it  (Figure  1).  Elevation  values  in  the 
basin  vary  from  411  meters  above  sea 
level  in  the  south-west  to  1615  meters 
in  the  south-east. 


284 


Figure  1. — Hill-shaded  view  of  the  DEM  covering  the  Lower 

Creek  Basin  test  area. 

In  deriving  a  landslide  susceptibility  index  map,  the  slope  gradient  was  first 
calculated  from  the  DEM  for  each  cell  and  then  reclassified  into  3  categories — 
greater  than  20°  representing  'high'  slopes,  10°-20°  for  'moderate',  and  <  10°  for 
'low'.  A  geological  map  was  also  created  with  the  same  grid  cell  size  and  geo- 
referencing  as  the  DEM.  Bedrock  was  classified  by  strength,  with  young  lava 
flows  and  intrusive  bodies  being  graded  as  'strong';  intermediate  age  clastic  rocks 
including  well-welded  ashflow  units  as  'moderate';  and  old  hydrothermally  altered 
clastic  rocks  (18-25  million  years  BP)  as  'weak'.  Using  the  Arc  GRID  software 
throughout  the  research,  the  slope  gradient  and  rock  strength  grids  were  combined 
to  yield  a  matrix  of  nine  values,  which  were  then  reclassified  into  three  levels  of 
landslide  susceptibility — 'high',  'moderate',  and  'low'  (see  Table  1).  This  is  the 
traditional  manner  in  which  GIS  is  used  to  solve  the  problem,  as  shown  in  Figure 
2,  and  the  combination  of  the  DEM  with  the  derived  slope  grid  and  geology  grid 
produces  a  single  susceptibility  index  grid. 

Clearly,  there  is  considerable  potential  here  for  applying  the  uncertainty  model 
to  assess  the  cumulative  effect  upon  the  final  susceptibility  index  derived  for  each 
cell — given  that  it  is  a  function  of  the  DEM  resolution  and  its  estimated  elevation 
error;  the  slope  gradient  calculation  for  each  cell;  and  the  method  of  reclassifying 
the  slope  values  and  rock  types  to  derive  the  index.  In  the  alternative  approach 
which  permits  uncertainty  to  be  assessed  (Figure  2),  elevation  noise  files  with 
varying  levels  of  spatial  autocorrelation  were  applied  to  the  original  DEM  to 
establish  corresponding  sets  of  slope  files  for  the  test  site.  The  range  of  ten  default 
values  of  p  as  described  in  the  previous  section  was  used.  The  realized  slope  files 
were  then  taken  in  turn  with  the  geology  file  and  used  to  derive  ten  susceptibility 
index  grids. 


285 


Traditional  Method 


Proposed  Method 


Noise  Realized  susceptibilty 

grids  grids 


Figure  2. — Comparison  of  the  traditional  and  proposed 
methods  of  calculating  landslide  susceptibility. 

Each  realized  grid  was  subtracted  from  the  original  susceptibility  grid  to  give  a 
set  of  ten  difference  grids  (one  for  each  value  of  p),  and  two  graphs  were  produced 
showing  the  mean  and  standard  deviation  of  the  differences  respectively  for  each 
grid  as  plotted  against  p.  Both  graphs  indicated  that  mean  and  standard  deviation 
values  increased  slightly  from  p  =  0  to  p  =  0.21,  then  dropped  suddenly  as  p 
approached  0.25.  The  value  of  p  =  0.21  was  estimated  as  the  point  at  which 
distortion  of  the  DEM  produced  an  output  that  was  both  stochastic  and  'truthful', 
however  this  aspect  of  the  uncertainty  model  remains  subjective  and  is  the  subject 
of  continuing  research. 

It  should  be  noted  here  that  the  actual  mean  and  standard  deviation  values  for 
each  realized  grid  have  little  significance  since  they  apply  to  nominal  data,  even 
though  the  susceptibility  indices  of  'low',  'moderate'  and  'high'  were  assigned 
numeric  values  to  enable  differencing.  Nevertheless,  the  graphs  still  serve  a  useful 
purpose  in  identifying  a  p  value  which  can  be  used  for  more  detailed  processing. 

The  procedure  was  then  repeated  by  taking  nine  elevation  error  grids  at  the 
single  value  of  p  =  0.21  and  creating  nine  realized  susceptibility  index  grids.  Since 
cells  exhibiting  high  instability  are  of  greatest  interest,  both  the  original 


286 


susceptibility  index  grid  and  the  realized  grids  were  reclassified  so  that  cells  with 
'low'  and  'moderate'  indices  were  assigned  a  value  of  zero,  while  cells  with  a 
'high'  index  were  allocated  a  value  of  unity.  The  nine  realized  grids  were  added 
together  to  produce  a  single  grid  containing  cumulative  values  denoting  the  number 
of  times  each  cell  scored  a  'high'  index  on  a  scale  of  0-9.  While  only  nine 
realizations  of  the  landslide  susceptibility  index  grid  were  created  due  to  time 
constraints  in  preparing  this  paper,  it  is  considered  that  a  much  greater  number 
(such  as  100  realizations)  would  be  used  in  practice  to  ensure  the  absence  of  any 
bias  in  the  elevation  error  files  used  to  perturb  the  DEM.  However,  the  results 
obtained  still  enable  useful  comments  to  be  made  about  the  process. 

To  visualize  the  results,  we  took  the  original  susceptibility  map  and  overlaid  all 
grid  cells  with  a  'high'  index  (shaded  black)  on  a  set  of  site  contours  produced  at 
100  m  interval  (Figure  3,  top).  For  comparison,  we  then  took  the  cumulative  index 
grid  described  previously  and  selected  all  cells  that  scored  a  'high'  index  at  least 
once  (Figure  3,  middle),  together  with  those  that  achieved  the  maximum  score  of 
nine  (Figure  3,  bottom).  These  cells  were  also  shaded  black  and  overlaid  on  the 
same  set  of  contours.  Alternatively,  we  could  choose  cells  on  the  basis  of  some 
probabilistic  measure,  for  example,  by  selecting  all  cells  which  score  a  'high'  value 
at  least  seven  out  of  nine  times. 


Table  2. — Relationship  between  'high' 

cell  scores  and 

area  of  the 

site  affected. 

'High  Score' 

No.  Cells 

Area  of  Site 

>1 

20,280 

29% 

>2 

16,843 

24% 

>3 

14,428 

21% 

>4 

12,549 

18% 

>5 

11,034 

16% 

>6 

9,651 

14% 

>7 

8,349 

12% 

>8 

6,872 

10% 

9 

4,982 

7% 

Original  Map 

10,168 

14% 

In  the  original  susceptibility  map, 
the  Boolean  nature  of  the  slope 
reclassification  into  categories  with 
thresholds  at  10°  and  20°  meant  that  a 
cell  could  receive  only  one  index 
value.  However,  with  the  proposed 
method  even  though  we  maintain  the 
same  thresholds  (and  their  values  are 
not  in  question  here),  we  can  see  how 
the  susceptibility  index  for  a  cell 
behaves  under  variation  of  the  input 
DEM.  Obviously,  cells  lying  on 
'weak'  or  'moderate'  rock  and  having 
slope  gradients  close  to  either  of  the 
thresholds  are  most  likely  to  exhibit  variation  in  their  susceptibility  index. 

Table  2  compares  the  number  of  'high'  cells  in  the  original  susceptibility  grid 
and  the  area  of  test  site  affected,  with  the  number  of  cells  achieving  a  'high'  score 
between  1-9  in  the  cumulative  realized  grid.  From  a  user's  perspective,  selecting 
cells  that  achieve  a  'high'  score  at  least  once  is  clearly  the  least  conservative  choice, 
given  that  the  elevations  of  many  of  these  cells  will  have  been  perturbed  several 
standard  deviations  from  the  mean  during  the  process.  We  believe  this  would  lead 
to  overestimation  of  the  area  affected.  On  the  other  hand,  selecting  only  those  cells 
that  attain  a  'high'  score  the  maximum  nine  times  is  the  most  conservative  approach 
and  most  likely  underestimates  the  true  area  of  the  site  susceptible  to  landslides. 
There  is,  however,  no  right  or  wrong  answer  to  the  selection  problem,  and  the  level 
of  risk  a  user  is  prepared  to  accept  in  the  end  product  becomes  a  matter  of  personal 
choice.  In  other  words,  users  must  learn  to  live  with  some  degree  of  uncertainty  in 
their  outputs  and  make  value  judgements  appropriate  to  individual  project 
requirements. 


287 


Figure  3. — (Top)  All  cells  in  the  original  susceptibility  index  map  scoring 
a  'high'  value  are  shown  in  black  overlaid  on  100m  contours.  (Middle)  All 
cells  shown  scoring  at  least  one  'high'  value  in  the  realization  process. 
(Bottom)  All  cells  shown  scoring  the  maximum  nine  'high'  value. 


288 


Finally,  there  is  another  error  source  which  will  affect  the  susceptibility  index 
that  has  not  been  considered  here — positional  uncertainty  in  the  geological 
boundaries  used  to  derive  rock  strengths.  At  this  stage,  the  uncertainty  model 
described  here  is  not  capable  of  handling  this  issue  since  a  solution  requires 
perturbation  of  the  original  geological  polygons  in  the  horizontal  plane.  However, 
research  is  currently  being  conducted  in  this  area  and  another  paper  elsewhere  in 
this  proceedings  describes  the  experimental  work  that  has  been  done  to  date  to 
overcome  the  problem  (see  Hunter  et  al.  1996). 

CONCLUSIONS 

In  this  paper,  the  authors  have  described  a  process  whereby  a  simple  slope 
stability  model  used  in  identifying  landslides  with  the  aid  of  GIS,  has  been 
modified  to  assess  the  uncertainty  in  the  final  susceptibility  map  as  a  result  of 
elevation  error  in  the  original  DEM  and  its  propagation  effects  upon  the  intermediate 
processes  involved.  Several  different,  but  equally  probable,  versions  of  the  input 
DEM  were  created  through  the  addition  of  random,  spatially  autocorrelated  noise 
(error)  files,  and  then  processed  to  produce  a  family  of  slope  stability  maps.  Users 
may  then  choose  the  level  of  risk  they  are  prepared  to  accept  in  the  final  output  and 
identify  cells  susceptible  to  landslides  that  meet  their  criteria. 

There  are  two  benefits  to  the  research.  Firstly,  the  inclusion  of  uncertainty 
assessment  in  the  process  has  the  potential  to  facilitate  improved  natural  resource 
management  decisions  in  relation  to  harvesting  and  subsequent  slope  stability.  And 
secondly,  from  a  user's  perspective  the  application  of  the  uncertainty  model  to 
applications  such  as  the  one  described  here  provides  a  new  dimension  to  their 
ability  to  understand  the  meaning  of  their  GIS  products. 

REFERENCES 

Cliff,  A.D.  &  Ord,  J.K.  1981.  Spatial  Processes:  Models  and  Applications,  Pion, 
London. 

Dyrness,  C.T.  1967.  "Mass  Soil  Movements  in  the  H.J.  Andrews  Experimental 

Forest".   Research  Paper  PNW-42,  Portland,  OR:  USD  A  Forest  Service. 

Pacific  Northwest  Forest  and  Range  Experiment  Station.  12p. 
Goodchild,  M.F.,  Guoqing,  S.  &  Shiren,  Y.  1992.  "Development  and  Test  of  an 

Error  Model  for  Categorical  Data",  International  Journal  of  Geographical 

Information  Systems,  Vol.  6,  No.  2,  pp.  87-104. 
Hunter,  G.J  &  Goodchild,  M.F.  1995.  "A  Methodology  for  Reporting  Uncertainty 

in  Spatial  Database  Products".   Urban  and  Regional  Information  Systems 

Journal,  vol.  7,  no.  ,  pp.  11-21. 
Hunter,  G.J.,  Hock,  B.  &  Robey,  M.  1996.  "Experimental  Development  of  a 

Model  of  Vector  Data  Uncertainty".  Proceedings  of  the  2nd  International 

Symposium  on  Spatial  Accuracy  Assessment  in  Natural  Resources  and 

Environmental  Modeling,  Fort  Collins,  Colorado,  8  pp. 
Hunter,  G.J.  &  Goodchild,  M.F.  (in  review).  "Modeling  the  Uncertainty  of  Slope 

Gradient  and  Aspect  Estimates  in  Spatial  Databases".  Submitted  to 

Geographical  Analysis,  22  pp. 
McKean,  J.,  Buechel,  S.  &  Gaydos,  L.  1991.  "Remote  Sensing  and  Landslide 

Hazard  Assessment".  Photo grammetric  Engineering  &  Remote  Sensing,  nol. 

57,  no.  9,  pp.  1185-1193. 


289 


Skaugset,  A.  1992.  "Slope  Stability  and  General  Hydrology  Research".  Forest 
Soils  and  Riparian  Zone  Management:  The  Contributions  of  Dr.  Henry  A. 
Froehlich  to  Forestry,  pp  23-45. 

Swanson,  F.J.,  &  Dyrness,  C.T.  1975.  "Impact  of  Clearcutting  and  Road 
Construction  on  Soil  Erosion  by  Landslides  in  the  Western  Cascade  Range, 
Oregon".  Geology,  vol  3,  pp.  393-396. 

Swanson  F.J.  &  Lienkaemper,  G.W.  1985.  "Geologic  Zoning  of  Slope 
Movements  in  Western  Oregon,  USA".  Proceedings  IVth  International 
Conference  and  Field  Workshop  on  Landslides,  pp.  41-45. 

Swanson,  F.J.,  Lienkaemper,  G.  &  Nakamura,  F.  (unpublished).  "Change  in 
Landslide  Hazard  through  Time  -  A  Case  Study  of  Forestry  Land  Use  and 
Climate  Change  in  the  Cascade  Mountains,  Oregon,  USA".  15pp. 

Swanson,  F.J.,  Swanson,  M.M.  &  Woods,  C.  1981.  "Analysis  of  Debris- 
Avalanche  Erosion  in  Steep  Forest  Lands:  An  Example  from  Mapleton,  Oregon, 
USA".  Proceedings  of  a  Symposium  on  Erosion  and  Sediment  Transport  in 
Pacific  Rim  Steeplands,  Davies,  T.R.H.,  Pearce,  A.J.  eds.,  Christchurch,  NZ, 
IAHS-AISH  Publication  No.  132,  pp.  67-75. 

BIOGRAPHICAL  SKETCHES 

Michelle  Murillo  is  a  graduate  student  with  the  Department  of  Forest  Science  at 
Oregon  State  University,  Corvallis.  She  gained  her  BS  from  the  University  of  New 
Mexico  in  Biology. 

Dr  Gary  Hunter  is  an  Assistant  Professor  with  the  Department  of  Geomatics 
and  Acting-Director  of  the  Center  for  Geographic  Information  Systems  and 
Modeling,  at  the  University  of  Melbourne,  Australia.  He  gained  his  PhD  from  that 
university  on  the  subject  of  handling  uncertainty  in  spatial  databases  and  is  the 
winner,  with  co-author  Professor  Michael  Goodchild,  of  the  Gerald  McCalden 
Prize  at  the  AURISA  94  conference  and  the  Horwood  Critique  Prize  at  the  URISA 
'93  and  '95  conferences,  for  papers  dealing  with  this  subject.  He  is  President  of 
the  Australasian  Urban  and  Regional  Information  Systems  Association  (AURISA) 
in  1996. 


290 


An  Accuracy  Assessment  of  a  Spatial 
Bioclimatic  Model 

D.W.  McKenney1,  B.G.  Mackey2,  M.F.  Hutchinson3,  R.A.  Sims4 

Abstract.— Spatially  explicit  models  of  forest  ecosystems  require  the 
integration  of  biophysical  data  at  a  range  of  scales.  Prediction  of  biotic 
response  should  be  based  on  spatially  explicit  models  of  the  driving 
environmental  processes,  in  particular  those  related  to  climate  and 
topography.  Ongoing  research  in  the  forests  of  Ontario  have  focused  on 
the  development  of  suitably  scaled  digital  elevation  models  and  meso- 
scaled  models  of  climatic  averages.  These  in  turn  have  been  used  to 
spatially  extend  various  biological  survey  data.  In  this  paper  we  present 
some  accuracy  assessments  of  the  spatial  climate  models  and  subsequent 
differences  in  species'  predicted  distributions. 


INTRODUCTION 

Resource  managers  require  reliable  spatial  information  about  species' 
distributions  to  help  make  decisions  such  as  establishing  monitoring  sites  or 
protected  area  networks.  How  important  a  particular  landscape  is  to  a  species  often 
depends  on  where  else  it  does  or  could  occur.  Spatial  information  can  help  provide 
context.  However  considerable  effort  is  required  to  develop  these  data.  Typical 
resource  inventories  such  as  aerial  photographs  or  remotely  sensed  data  provide 
relatively  little  information  about  non-commercial  and/or  non-overstory  species, 
so  some  other  type  of  modelling  is  required. 

Theory  would  suggest  that  ecological  response,  such  as  a  species'  distribution, 
is  a  function  of  a  complex  myriad  of  factors  including  disturbance  regimes,  human 
inputs  (eg.  land  use  history),  biological  regimes  (eg.  interspecies  competition)  and 
a  number  of  physical  determinants.  The  latter  include  thermal,  radiation,  moisture 
and  nutrient  regimes.  Data  on  climate,  terrain  and  soil  parent  material  are  the 
basic  inputs  to  modelling  these  regimes  across  landscapes.  If  spatial  data  on  the 


Senior  Economist,  Canadian  Forest  Service  -  Sault  Ste.  Mane,  Ontario,  Canada. 

Senior  Lecturer,  Department  of  Geography,  The  Australian  National  University,  Canberra,  Australia. 

Senior  Fellow,  Centre  for  Resource  and  Environmental  Studies,  The  Australian  National  University,  Canberra,  Australia. 

Research  Scientist,  Canadian  Forest  Service  -  Sault  Ste.  Mane,  Ontario,  Canada. 


291 


regimes  are  available  and  species'  responses  can  be  calibrated  with  the  regimes, 
then  there  becomes  an  enhanced  capability  to  extrapolate  information  gained  from 
analysis  of  field  based  observations.  This  type  of  empirical  framework  is  directly 
relevant  to  the  problem  of  spatially  quantifying  species'  distributions,  abundance, 
and  productivity  (eg.  Mackey  1994).  In  this  paper,  some  accuracy  issues 
associated  with  a  spatial  climate  model  and  the  problem  of  extrapolating  site  data 
to  quantify  a  species'  (jack  pine,  Pinus  banksiana)  potential  distribution  are 
investigated. 

A  SPATIAL  CLIMATE  MODEL  FOR  THE  PROVINCE  OF  ONTARIO 

It  is  well  accepted  that  climate  exerts  a  strong  control  on  both  the  distribution 
and  productivity  of  plant  species.  However,  correlations  between  various  climate 
parameters  and  the  ecological  response  of  interest  is  often  poor.  One  reason  for 
this  is  that  weather  stations  are  located  well  away  from  field  survey  points  and 
any  estimates  of  climatic  parameters  involve  some  interpolation.  Some  relatively 
new  methods  for  interpolating  climate  provide  opportunities  for  generating  more 
reliable  spatial  estimates  of  climatic  parameters  ex  situ  than  previously  possible 
(see  Hutchinson  1995,  1987).  Applications  of  these  methods  to  estimate  long  term 
monthly  mean  climate  values  at  points  of  survey  and  subsequently  explain  spatial 
variation  in  ecological  response  have  been  successful  (eg.  Nix  1986,  Mackey 
1994,  Yee  and  Mitchell  1991).  The  methods  make  use  of  partial  thin  plate 
smoothing  splines  (Eq.  1)  to  interpolate  historical  weather  station  data: 

p 

j 

where  /  is  an  unknown  smooth  to  be  estimated,  are  a  set  of  p  known  functions 
and  the  Pj  are  a  set  of  unknown  parameters  which  have  to  be  estimated.  The  Xj 
generally  represent  coordinates  in  two  or  three  dimensional  space.  The  e{  are  zero 
mean  random  errors.  Hutchinson  (1995)  describes  the  estimation  procedures  in 
detail.  In  applications  to  date,  climate  surfaces  have  been  developed  as  a  function 
of  latitude,  longitude  and  elevation  (eg.  Nix  1986,  Mitchell  1991,  Mackey  et  al. 
1996)  although  any  number  of  independent  variables  are  possible  in  principal  (eg. 
distance  from  a  major  water  body).  These  three  parameters  are  known  to  be 
important  determinants  of  long-term  climate.  The  resultant  climate  surfaces  can 
also  be  mapped  relatively  easily  in  a  GIS  using  software  which  couples  the 
functions  to  a  Digital  Elevation  Model  (DEM)  (see  Hutchinson  1993).  A  DEM  is 
a  regular  grid  of  latitude,  longitude  and  elevation. 

The  main  advantage  of  the  thin  plate  spline  techniques  over  competing 
methods  is  that  they  do  not  require  any  a  priori  estimation  of  the  spatial  auto- 
covariance  structure.  In  addition,  data  smoothing  can  be  optimized  by  minimizing 
the  generalized  cross  validation  statistic  (Hutchinson  1995,  Hutchinson  and  Gessler 


292 


1994). 

Mackey  et  al.  (1996  in  press)  used  these  methods  to  generate  climate  surfaces 
for  the  province  of  Ontario  as  a  function  of  x,  y  (latitude  and  longitude)  and  z  (km 
above  sea  level).  Table  1  summarizes  the  standard  errors  of  the  basic  surfaces  of 
monthly  mean  maximum  temperature,  monthly  mean  minimum  temperature  and 
monthly  precipitation.  Besides  mapping  the  surfaces  and  secondary  indices  in  a 
GIS  by  coupling  the  functions  to  a  DEM,  estimates  of  climatic  values  can  be 
generated  at  field  survey  locations.  The  latitude,  longitude  and  elevation  (x,  y  and 
z)  of  each  survey  location  are  required.  Hence,  estimates  of  various  climate 
variables  can  be  estimated  and  climate/plant  or  plant/animal  relations  can  be 
subsequently  examined.  Spatial  predictions  can  be  made  by  linking  the  statistical 
functions  to  spatial  databases  of  the  independent  variables. 


Table  1.— Number  of  data  points  and  approximate  standard  errors  of  fitted  climate 
surfaces  for  Ontario. 


Climate  Variable 

Number  of  Data  Points 

Standard  Error 

monthly  mean  max.  temp. 

471 

0.13  to  0.26°C 

monthly  mean  min.temp. 

471 

0.27  to  0.42°C 

monthly  precipitation 
(old  surface) 

471 

3.2  to  7.6% 

monthly  precipitation 

470 

3.3  to  8.2% 

(new  surface) 

(less  Ignace) 

Climate  Estimate  Errors  Associated  With  Incorrect  Field  Survey  Locations 

Considerable  effort  has  been  placed  on  the  collation  of  existing  biological 
survey  data  in  Ontario  so  that  these  methods  can  be  applied.  For  example,  there 
are  now  over  4100  forest  ecosystem  classification  (FEC)  plots  in  existence 
throughout  the  province  (Sims  and  Uhlig  1992).  These  data  are  essentially 
vegetation  surveys  in  which  numerous  characteristics  were  measured  in  10  by  10 
metre  plots  and  included  understorey  and  overstory  plants,  soil  properties  and 
some  mensurational  attributes.  Some  of  the  newer  plots  used  GPS  technology  to 
record  location  data.  If  location  information  was  not  already  recorded,  the  original 
survey  records  and  topographic  maps  were  used  to  record  x,  y  and  z  (it  is  also 
possible  to  use  a  DEM  to  derive  estimates  of  elevation  if  only  x  and  y  are 
available). 

The  importance  of  accurate  location  information  will  vary  spatially.  For 
example;  steeper  climatic  gradients  will  exist  where  elevation  changes  are  more 
abrupt,  or  near  very  large  water  bodies  such  as  the  Great  Lakes.  To  gauge  the 
importance  of  location  information  across  the  province,  an  analysis  of  the  location 
data  associated  with  observations  of  jack  pine  was  performed.  Of  the  4100  FEC 


293 


plots,  951  contained  jack  pine.  The  x,y,z  coordinates  of  each  of  these  were 
perturbed  in  two  ways.  Figure  1  reports  an  example  where  the  elevation  (z)  values 
were  increased  by  50  metres  (elevations  above  sea  level  in  Ontario  range  from  0 
metres  near  Hudson  Bay  to  over  610  metres  northwest  of  Lake  Superior).  New 
climate  values  were  generated  with  these  increased  elevation  values  and  compared 
to  the  original  estimates.  Results  for  three  climatic  values  (growing  season  length, 
maximum  temperature  of  the  hottest  month  and  precipitation  of  the  coldest 
quarter)  are  shown  in  figure  1.  The  original  range  of  values  at  the  951  plots  is  also 
reported  in  the  figure. 

In  addition  to  the  elevation  perturbations,  x  and  y  coordinates  were 
systematically  altered.  Figure  1  reports  the  results  of  increasing  x  and  y  by  six 
minutes.  Across  the  range  of  the  plot  locations,  the  implication  in  terms  of 
distance  is  approximately  13  km  northwest  of  each  "true"  geographic  plot  position. 

The  changes  in  elevation  and  latitude/longitude  produced  similar  changes  in 
growing  season  length.  The  changes  in  temperature  and  precipitation  were 
consistent  with  the  relative  scaling  of  the  input  data.  The  differences  from  the 
original  values  are  marginal.  The  biggest  impact  was  on  precipitation  in  the 
coldest  quarter.  In  this  case  the  elevation  change  caused  a  general  decrease  in 
values  across  all  sites  (from  -1.5%  to  -5.7%  in  relation  to  the  original  values).  The 
lat/long  shift  resulted  in  a  -6.9%  to  9.9%  change  in  relation  to  the  original  values. 
These  differences  are  for  the  most  part  within  the  standard  errors  reported  in  table 
1.  The  results  illustrate  the  spatially  varying  dependency  of  climate  on  these  3 
variables. 

SPATIAL  PREDICTIONS  OF  THE  RANGE  OF  JACK  PINE  IN 

ONTARIO 

As  mentioned,  the  ability  to  generate  grids  of  climatic  values  and  estimates  of 
climate  at  points  of  survey  offer  opportunities  to  develop  spatial  predictions  of 
animal-plant/climate  relations.  A  variety  of  approaches  are  possible  depending  on 
the  nature  of  the  biological  site  data  including  the  use  of  statistical  relationships 
(eg.  presence  of  an  organism  =  /(minimum  temperature  of  the  coldest  month).  An 
approach  that  has  proven  useful  when  only  presence  data  is  available  is  a  grid 
matching  procedure  (Nix,  1986).  In  this  case  a  climatic  profile  of  the  species  is 
generated,  based  on  observations  of  where  it  occurs  and  the  estimates  of  the 
climatic  values  at  each  of  those  locations.  This  profile  is  then  matched  to  grids  of 
those  climatic  parameters.  This  is  the  BIOCLIM/BIOMAP  procedure  developed 
by  Nix  (1986)  and  colleagues  at  the  Australian  National  University  (McMahon  et 
al.  1995). 

BIOCLIM  makes  use  of  climate  surfaces,  DEMs  and  species'  location 
information  to  produce  bioclimatic  profiles.  These  profiles  essentially  describe  the 
climatic  conditions  (eg.  the  range  of  mean  annual  temperatures,  annual 
precipitation)  sampled  by  a  plot  network.  BIOMAP  takes  the  BIOCLIM  output 


294 


a 


b 


800  - 


600  - 


400  - 


200  - 


0 


800  - 


600  - 


400 


200  - 


-1    0     1     2     3     4     5  -1  012345 

Growing  season  length  (range  of  original  values  =  1 62  to  21 1  days) 


800  - 


-0.6    -0.4    -0.2     0.0     0.2     0.4  -0.6    -0.4    -0.2     0.0     0.2  0.4 

Maximum  temperature  of  the  hottest  month  (range  of  original  values  =  19.8  to  26.5°C) 


800  - 


600  - 


400 


200 


800  - 


600  - 


400  - 


200  - 


0 

-12     -8     -4      0      4      8      12  -12     -8     -4      0      4      8  12 

Precipitation  of  the  coldest  quarter  (range  of  original  values  =  69  to  272  mm) 

Figure  1.  Differences  in  three  climate  variable  estimates  at  951  FEC  plots  when 
perturbing  latitude/longitude  +6  minutes  (column  a)  and  elevation  +50  metres  (column  b). 
The  range  of  differences  as  a  percentage  of  original  values  is  presented  for  each  case  (%  diff). 


295 


and  identifies  (flags)  grid  cells  which  match  the  bioclimate  profile  for  the  selected 
climatic  variables.  The  resulting  data  can  be  input  into  a  GIS  and  interpreted  as 
a  spatial  prediction  of  the  climatic  domain  of  the  species.  In  this  way,  landscapes 
can  be  identified  which  are  climatically  suitable  for  the  species.  Whether  the 
species  actually  occurs  will  depend  on  the  influence  of  other  environmental 
processes. 

This  approach  to  predicting  the  potential  domain  of  a  species  is  relatively 
transparent  to  interpretation.  However  there  are  limitations;  all  of  the  selected 
climatic  variables  used  in  the  BIOMAP  predictions  are  assumed  to  have  an  equal 
impact  on  the  species'  distribution.  Statistical  options  are  limited  when  the 
available  data  indicate  only  where  a  species  is  present;  ie.  there  are  no 
observations  of  where  it  is  absent.  BIOCLIM/BIOMAP  is  a  repeatable  and  useful 
method  for  analyzing  these  types  of  survey  data  and  can  be  easily  updated  with 
new  observations.  However  the  spatial  prediction  is  also  contingent  on  the 
representativeness  of  the  underlying  biological  site  data  and  quality  of  the  climate 
surfaces. 

Figure  2a  is  a  BIOMAP  of  jack  pine  based  on  analysis  of  951  province- wide 
FEC  plots.  Future  work  by  the  authors  will  provide  a  detailed  discussion  of  this 
type  of  analysis  for  a  suite  of  commercial  tree  species  in  Ontario).  A  set  of 
bioclimate  parameters  (eg.  annual  mean  temperature,  precipitation  in  the  warmest 
and  coldest  quarters)  derived  from  the  monthly  mean  climate  variables  were 
coupled  to  a  1 -kilometre  DEM  of  the  province  in  order  to  produce  the  BIOMAP. 
The  resulting  dark  areas  represent  the  "core"  climate  domain  for  the  species.  Grid 
cells  are  identified  as  a  core  area  if  all  of  their  climatic  parameter  values  fall 
within  10-90%  of  the  survey  data  climatic  values  (ie.  the  10-90  percentiles  of  the 
survey  climatic  parameters  are  the  climatic  conditions  where  most  of  the  site  data 
occur).  The  lightest  shade  of  grey  indicates  those  grid  cells  that  fall  within  the 
range  of  climatic  values  defined  by  the  survey  data,  indicating  more  marginal 
conditions. 

Noted  in  figure  2a  is  a  "hole"  near  Ignace,  Ontario,  which  suggests  the  area 
is  outside  the  climatic  domain  of  jack  pine.  This  outcome  has  been  noticed  with 
BIOMAP  analysis  of  several  plants  and  animals  and  has  been  a  source  of  some 
interest  given  an  a  priori  expectation  that  the  region  would  be  quite  suitable  to 
many  of  the  species.  During  an  investigation  of  the  spatial  pattern  of  the 
bioclimatic  parameters  in  that  region,  it  was  found  that  precipitation  during  the 
coldest  quarter  is  significantly  drier  there  than  the  surrounding  area.  Figure  3a 
shows  the  precipitation  surface  for  that  area  in  close  detail  and  also  identifies  the 
locations  of  FEC  plots  in  that  part  of  the  province.  Clearly  there  are  no  FEC  plots 
in  this  area.  Thus,  being  an  empirical  approach,  the  BIOMAP  output  is  doing  what 
would  be  expected;  producing  a  "hole"  at  those  grid  cells  where  at  least  one  of  the 
climate  parameters  has  values  outside  the  range  of  the  climatic  values  for  known 
occurrences  of  the  species.  An  implication  is  that  the  area  is  outside  the  range  of 
jack  pine. 


296 


Figure  2.  Climatic  domain  analysis  for  jack  pine  in  Ontario  using  original  precipitation 
surface  (a)  and  new  precipitation  surface  (b) .  Map  (c)  is  the  difference  between  (a)  and  (b). 


297 


One  question  is  whether  the  underlying  climate  surface  is  robust  (see  figure 
3a).  However,  the  apparent  anomaly  is  a  reflection  of  the  input  data.  It  was 
determined  that  one  weather  station  (Ignace),  at  the  centre  of  the  "hole"  was 
reporting  much  drier  conditions  than  nearby  stations.  For  example,  long-term 
November  to  April  precipitation  is  138mm  compared  to  17 1-2 12mm  for  nearby 
stations.  As  well  as  having  a  shorter  history  of  record  collection  than  other 
stations,  further  investigations  revealed  that  Canadian  Atmospheric  Environment 
Service  experts  had  several  reasons  to  be  suspicious  of  the  quality  of  data  from 
this  particular  station.  In  fact,  this  Ignace  station  is  now  closed. 

Given  this,  a  new  precipitation  surface  was  created  without  the  Ignace  station 
data.  Results  are  shown  in  figure  3b.  The  steep  gradient  associated  with 
precipitation  of  the  coldest  quarter  has  been  smoothed.  The  standard  error  of  the 
monthly  precipitation  surfaces  using  data  from  470  weather  stations  ranged  from 
3.3%  to  8.2%  compared  to  3.2%  to  7.6%  (see  table  1).  The  implication  in  terms 
of  the  predicted  climatic  range  of  jack  pine  is  shown  in  figure  2b.  The  "Ignace 
hole"  shown  in  figure  3a  is  now  part  of  the  core  climatic  range  for  jack  pine.  The 
difference  between  figures  2a  and  2b  is  shown  in  figure  2c.  Although  there  are 
some  changes  across  the  province,  the  most  significant  changes  occurred  in  the 
Ignace  area  where  grid  cells  that  were  flagged  as  outside  the  species  range  are 
now  part  of  its  core  range. 

CONCLUDING  COMMENTS 

Reliable,  quantitative  spatial  data  should  ideally  be  used  to  help  make 
assessments  about  the  potential  of  landscapes  to  support  particular  species.  Species 
distributions  must  be  modelled  through  integration  of  various  data  types  and 
sensitivity  analysis  of  the  results  to  input  values  should  occur.  Empirical  models 
are  generally  driven  by  the  input  data,  in  this  paper  these  are  the  climate  and 
biological  data.  The  results  presented  here  reiterate  one  truism  well  known  to 
ecological  researchers:  data  quality  control  should  be  a  paramount  concern. 
However  it  is  clear  that  some  data  quality  issues  will  not  arise  without  extensive 
investigation.  Prior  to  the  analysis  of  site  data  in  conjunction  with  climate,  we 
were  quite  satisfied  with  the  Ontario  climate  model.  In  this  case  the  quest  to 
derive  spatial  information  from  site-based  observations  revealed  an  area  of  Ontario 
that  lacked  FEC  plots  and  had  questionable  weather  station  data. 

From  a  narrower  perspective,  the  analysis  summarized  in  figure  1  suggests 
considerable  scope  to  make  use  of  historical  field  data  in  Ontario.  There  are 
numerous  vegetation  and  wildlife  survey  datasets  available  in  Ontario  to  which 
latitude,  longitude  and  elevation  could  be  appended.  The  analysis  of  differences 
in  climate  estimates  in  this  paper  should  allay  concerns  about  moderate  errors  in 
location  records. 


298 


a  )  I      I  60-70mm 

90- 100mm 


I      i  70-80mm 
100- 11  Omm 


I      I  80-90mm 
11 0-1 20mm 


Figure  3.  Precipitation  in  coldest  quarter  at  the  'Ignace  hole'  (48:45N--50:00N,  91:15W 
■92:45W)  with  nearby  FEC  plots  overlaid. 


299 


ACKNOWLEDGEMENTS 

The  authors  would  like  to  thank  Kathy  Campbell,  Janice  McKee,  Brian  Zavitz 
and  Yin-Quin  Yang  for  their  assistance  in  preparing  the  figures.  Funding  has  come 
from  various  sources  including  the  Canada- Ontario  Northern  Ontario  Development 
Agreement  Program,  the  Canadian  Forest  Service  and  Ontario's  Sustainable 
Forestry  Initiative,  Ontario  Forest  Research  Institute. 

REFERENCES 

Hutchinson,  M.F.,  1987,  Methods  of  generating  weather  variables:  in  Agricultural 

environments:  Characterisation,  classification  and  mapping,  A.H.  Bunting  ed., 

Wallingford  CAB  International:  149-157. 
Hutchinson,  M.F.,  1993,  Development  of  a  continent- wide  DEM  with  applications 

to  terrain  and  climate  analysis:  in  Environmental  modelling  with  GIS,  M.F. 

Goodchild,  B.O.  Parks,  and  L.T.  Stewart  eds.,  Ox.  Univ.  Press:  392-399. 
Hutchinson,  M.F.,  1995,  Interpolating  mean  rainfall  using  thin  plate  smoothing 

splines,  International  Journal  of  Geographical  Information  Systems,  (9):  385- 

403. 

Hutchinson  M.F.,  and  Gessler,  P.E.,  1994,  Splines  -  more  than  just  a  smooth 

interpolator,  Geoderma,  (62):  45-67. 
Mackey,  B.G,  1994,  Predicting  the  potential  distribution  of  rainforest  structural 

characteristics,  Journal  of  Vegetation  Science,  (5): 3-54. 
Mackey,  B.G.,  McKenney,  D.W.,  Widdifield,  C,  Sims,  R.A.,  Lawrence,  K., 

Szcyrek,  N.,  1994,  A  new  digital  elevation  model  of  Ontario,  Nat.  Resour. 

Canada  -  Canadian  Forest  Service  -  Sault  Ste.  Marie  ON,  NODA/NFP 

Technical  Report  TR-6:  26  p. 
Mackey,  B.G.,  McKenney,  D.W.,  Yang,  Y.,  McMahon,  J.P.,  Hutchinson,  M.F., 

1996,  Site  regions  revisited:  A  climatic  analysis  of  Hills'  site  regions  for  the 

province  of  Ontario  using  a  parametric  method,  Canadian  Journal  of  Forest 

Research,  (26):  333-354. 
McMahon,  J.P.,  Hutchinson,  M.F.,  Nix,  H.A.  and  Ord,  K.D.  1995.  ANUCLIM 

User's  Guide.  Draft  Report,  Centre  for  Resource  and  Environmental  Studies, 

Australian  National  University,  Canberra. 
Mitchell,  N.D.,  1991,  The  derivation  of  climate  surfaces  for  New  Zealand,  and 

their  application  to  the  bioclimatic  analysis  of  the  distribution  of  Kauri 

(Agathis  australis),  New  Zealand  Journal  of  Royal  Society,  (21):  13-24. 
Nix,  H.A.,  1986,  A  biogeographic  analysis  of  Australian  elapid  snakes:  in  Atlas 

of  elapid  snakes  of  Australia,  R.  Longmore  ed.,  Australian  flora  and  fauna 

series  No.  7.  Australian  Government  Publ.  Service,  Canberra  Australia:  4-15. 
Sims,  R.A.,  Uhlig,  P.,  1992,  The  current  status  of  forest  site  classifications  in 

Ontario,  Forestry  Chronicle,  (68):64~-77. 
Yee,  T.W.  and  Mitchell,  N.D.,  1991,  Generalized  additive  models  in  plant 

ecology,  Journal  of  Vegetation  Science,  (2):  587-602. 


300 


An  Assessment  of  the  Horizontal 
Accuracy  of  Interim  Terrain  Data 


L.A.  Fatale1,  J.A.  Messmore2,  and  J.R.  Ackeret3 


Abstract.— Within  Army  and  throughout  the  private  sector,  the  accuracy  of 
Digital  Topographic  Data  (DTD)  is  often  taken  for  granted.  Users  assume 
terrain  data  to  be  accurate  merely  because  of  its  digital  nature.  This  paper 
describes  a  study  to  determine  the  horizontal  accuracy  of  Interim  Terrain  Data 
(ITD),  a  1 : 5 0,000  scale  terrain  analysis  data  base.  Due  to  its  interim  nature, 
ITD  may  be  digitized  from  hard  copy  source  ("carto-controlled")  or  produced 
solely  from  photography  ("photo-controlled").  The  producer  (Defense 
Mapping  Agency,  DMA)  provides  no  definitive  accuracy  statement  with  the 
data.  Over  400  ITD  features  were  chosen  for  analysis.  Each  feature  was 
located  in  the  field  using  a  Precise  Lightweight  Global  Positioning  System 
Receiver  (PLGR)  in  the  Precise  Positioning  Service  (PPS)  mode  (-10  meter 
accuracy).  Universal  Transverse  Mercator  (UTM)  coordinates  obtained  from 
the  PLGR  were  annotated  for  each  feature.  UTM  coordinates  for  coincident 
features  in  the  ITD  digital  files  were  subsequently  extracted  by  GIS  query  in 
order  to  statistically  compare  the  digital  data  to  the  field  data,  calculate  the 
offsets,  and  thereby  determine  the  horizontal  accuracy  of  the  ITD.  Potential 
accuracy  differences  between  the  carto-controlled  and  photo-controlled  ITD 
and  the  variability  of  natural  vs.  man-made  features  are  also  discussed. 

INTRODUCTION  &  BACKGROUND 

Interim  Terrain  Data  (ITD)  was  first  produced  by  the  Defense  Mapping  Agency 
(DMA)  in  1987.  Production  of  this  data  was  initiated  to  fill  the  Army's  urgent  need 
for  a  digital  terrain  analysis  product  in  support  of  emerging  topographic  and 
Command  and  Control  (C2)  systems.  Earlier,  in  1984,  the  Army  had  documented 
a  digital  topographic  data  (DTD)  requirement  that  encompassed  all  Army's  known 
and  anticipated  tactical  DTD  requirements.  However,  production  against  this 
requirement,  described  as  Tactical  Terrain  Data  (TTD),  was  not  scheduled  to 
commence  until  1996.  In  order  to  get  similar  data  to  Army  users,  DMA  initiated 


1  Physical  Scientist,   U.S.  Army  Topographic  Engineering  Center,  Alexandria ,  VA 

2  Supervisory  Physical  Scientist ,   U.S.  Army  Topographic  Engineering  Center, 

Alexandria ,  VA 

3  Physical  Scientist,   U.S.  Army  Topographic  Engineering  Center,  Alexandria ,  VA 


301 


production  of  an  interim  product  called  ITD,  the  goal  of  which  was  to  fill  the  data 
void  until  TTD  production  came  on  line.  Due  to  the  interim  nature  of  this  data  set 
and  the  use  of  more  than  one  production  method,  DMA  does  not  state  a  formal 
horizontal  accuracy  measure  for  this  product. 

As  the  center  of  expertise  for  DTD  within  the  Army,  the  Digital  Concepts  and 
Analysis  Center  (DC AC)  at  the  U.S.  Army  Topographic  Engineering  Center  (TEC) 
formulated  this  study  to  characterize  the  horizontal  accuracy  of  ITD.  Knowledge  of 
horizontal  accuracy  is  important  to  the  Army  as  well  as  other  DOD  users  because  it 
affects  what  are  appropriate  uses  of  the  product.  For  example,  is  ITD  as  accurate 
as  the  Topographic  Line  Map  (TLM),  which  may  be  used  for  targeting  purposes,  or 
is  it  only  as  accurate  as  a  city  graphic  which  is  not  an  appropriate  source  for 
targeting?  Fortunately,  along  with  the  need  to  know  ITD's  horizontal  accuracy, 
technology  for  rapid  position  determination  has  become  available  in  the  form  of 
Global  Positioning  Systems  (GPS).  This  technology  was  applied  to  the 
determination  of  the  horizontal  positions  of  well-defined  ITD  features. 

The  ITD  is  produced  through  two  production  processes,  one  is  a  hard-copy 
digitizing  process,  and  the  other  is  a  pure  photogrammetric  production  process.  In 
the  hard-copy  process,  tactical  terrain  analysis  data  base  (TTADB)  mylar  overlays 
(which  are  geo-registered  to  the  corresponding  1 :50,000-scale  TLMs)  are  digitized, 
thereby  resulting  in  a  "carto-controlled"  ITD  product.  The  photogrammetric 
production  process  is  effected  using  imagery  on  DMAs  feature  extraction  (FE) 
workstation,  thereby  resulting  in  a  "photo-controlled"  ITD  product.  It  should  be 
noted  that  ITD  cells  from  both  of  these  widely  divergent  production  processes  are 
currently  distributed  and  in  broad  use.  Carto-controlled  cells  make  up  the  bulk  of 
currently  available  ITD  (»840  of  1340  cells).  The  -500  remaining  cells  are  photo- 
controlled. 

PURPOSE 

The  purpose  of  this  study  was  to  provide  the  Mapping  Charting  &  Geodesy 
(MC&G)  user  community  with  information  about  the  horizontal  accuracy  of  ITD 
features.  It  is  intended  to  give  the  user  a  level  of  confidence  in  regard  to  ITD 
accuracy  and  to  establish  the  utility  of  ITD  for  various  applications.  The  study  also 
identified  accuracy  differences  between  cartographic  and  photogrammetrically 
controlled  ITD. 

METHODOLOGY  &  FIELD  COLLECTION  PROCEDURE 

To  facilitate  DCAC's  ITD  accuracy  investigations,  ITD  hard-copy  plots  and 
digital  data  for  10  cells  in  the  continental  U.S.  (CONUS)  were  obtained  from  DMA 
along  with  the  associated  TLMs.  The  use  of  the  hard-copy  plots  (from  digital  ITD) 
instead  of  the  TTADB  ensured  that  features  identified  for  analysis  would  be 
represented  in  the  digital  data.  Each  of  these  plots  covered  approximately  a  15 -by 


302 


15 -minute  area.  There  were  four  photo-controlled  cells  in  the  Alamagordo/Las 
Cruces,  NM  area  as  well  as  six  carto-controlled  cells;  two  in  the  Ft.  Bragg,  NC,  area 
and  four  in  the  vicinity  of  Ft.  Hood,  TX. 

Scientists  at  DCAC  reviewed  the  hard-copy  plots  and  identified  appropriate 
features  for  field  reconnaissance.  Two  factors  weighed  heavily  in  the  selection  of 
features.  The  first  factor  was  the  existence  of  a  distinguishable  point  on  the  ground 
(e.g.,  a  road  intersection,  bridge,  dam,  or  fence  line).  The  second  factor  was 
accessibility  for  measurement  of  its  horizontal  placement.  No  vertical  accuracy 
criteria,  such  as  feature  height  or  height  above  mean  sea  level,  were  considered.  The 
selected  features  were  then  annotated  on  the  coinciding  1 : 50,000  TLM  for  field  use. 
They  comprised  a  representative  sample  from  the  Transportation  (TRANS),  Surface 
Drainage  (SD),  Obstacles  (OBS),  and  Vegetation  (VEG)  overlays  since  these 
contained  point  or  line  features  that  were  more  easily  recognizable  in  the  field. 
Features  in  the  Surface  Configuration  and  Soils  overlays  were  not  utilized  for  this 
study.  Although  important  for  many  applications,  these  features  are  not  easily 
distinguishable  on  the  ground  and  are  difficult  to  measure  accurately. 

Depending  on  the  density  of  the  data,  between  14  and  73  features  on  each  of  the 
10  cells  in  the  study  areas  were  chosen  for  analysis.  The  number  and  variety  of 
features  chosen  were  those  deemed  necessary  to  assure  representative  coverage 
throughout  each  cell.  Over  400  features  were  subsequently  visited  in  the  field.  Each 
site  was  located  using  a  Precise  Lightweight  Global  Positioning  System  Receiver 
(PLGR)  in  the  Precise  Positioning  Service  (PPS)  mode  (-10  meter  accuracy). 
Universal  Transverse  Mercator  (UTM)  coordinates  derived  from  the  PLGR  were 
annotated  for  each. 

ANALYSIS 
Data  Preparation 

The  digital  ITD  files  provided  by  DMA  were  partitioned  on  the  provided  media 
(9-track  tape/6250bpi)  by  individual  sheet  number  and  thematic  layer(s).  Related  file 
header  data  was  used  to  acquire  descriptive  information  (metadata)  pertaining  to  the 
individual  data  sets.  The  digital  ITD  files  were  archived  in  Standard  Linear  Format 
(SLF).  Arc/Info  (V.  7.0)  was  used  to  import  the  ITD-SLF  files  read  from  the  tape. 
Upon  import,  the  data  was  converted  into  Arc/Info  coverages  using  the  "SLF ARC" 
command.  Arc/Info's  GIS  functionalities  enabled  us  to: 

1 .  Identify  &  display  selected  features  by  thematic  layer,  feature  type, 
and  feature  attributes. 

2.  Precisely  extract  UTM  coordinates  for  selected  features. 


303 


The  Arc/Info  coverages  were  geo-referenced  in  WGS-84  spherical  geographic 
coordinates  (LAT-  LONG).  Each  coverage,  maintained  in  single-coordinate 
precision  (seven  significant  digits),  was  subsequently  projected  into  a  WGS-84  planar 
coordinate  system  (UTM)  which  was  consistent  with  the  GPS  field  collection 
coordinate  system.  Projection  of  coordinates  from  LAT-LON  to  real-world  UTM 
coordinates  via  the  Arc/Info  projection  utilities  was  determined  to  be  a  virtually  error 
free  process.  Since  both  input  and  output  coverages  were  stored  in  WGS-84  and 
required  no  additional  datum  transformations,  the  statistical  significance  of  the  root- 
mean-square  (RMS)  error  was  <  0.010.  [NOTE:  A  perfect  projection  with  an  RMS 
error  of  0.000  is  not  possible  with  real-world  data.]  Customized  graphical  user 
interfaces  were  developed  utilizing  Arc/Info's  Arc  Macro  Language  (AML)  to 
facilitate  extraction  of  digital  ITD  feature  coordinates. 

Digital  ITD  Coordinate  Extraction 

UTM  coordinates  for  selected  features  were  extracted  in  order  to  compare  the 
field  data  to  the  digital  data,  calculate  the  offsets,  and  thereby  determine  the 
horizontal  accuracy  of  ITD.  Thematic  coverages  for  each  cell  in  the  study  area  were 
displayed  with  the  features  appearing  as  unsymbolized  center-line  data.  Next,  those 
ground  features  visited  in  the  field  were  interactively  identified  in  the  digital  file.  A 
coordinate  pair  was  extracted  and  recorded  alongside  the  coordinate  pair  of  the 
coincident  ground  feature.  Finally,  a  series  of  statistical  analyses  were  performed  on 
both  sets  of  coordinates. 

Statistical  Analyses 

Coordinates  for  the  photo-controlled  and  the  carto-controlled  ITD  features  were 
stored  in  different  files  which  were  further  delineated  by  eastings  and  northings. 
These  files  were  then  grouped  by  specific  feature  types  (e.g.,  "surface  drainage 
features  from  photo  controlled  data").  Standard  statistical  measures  such  as  mean, 
standard  deviation,  95  percent  confidence  interval  and  a  t-statistic  with  an  associated 
critical  value  were  calculated  for  each  data  file.  The  mean  and  standard  deviation 
determine  the  size  of  the  error.  The  95  percent  confidence  interval  indicates  the 
potential  magnitude  of  that  error  in  future  measurements.  The  t-statistic  determines 
whether  there  is  a  significant  difference  between  two  populations  .  The  critical  value 
is  the  threshold  that  the  T-statistic  must  exceed,  at  a  given  confidence  level  and 
sample  size,  to  ensure  rejection  of  the  null  hypothesis  (i.e.,  for  this  study,  that  there 
is  no  difference  between  the  field  data  coordinates  and  those  of  the  ITD  data). 


304 


DISCUSSION/FINDINGS 


During  the  field  work,  DC  AC  scientists  visited  42 1  feature  locations,  collecting 
an  easting  and  a  northing  at  each  for  a  total  of  842  observations  (combined  photo- 
and  carto-controlled  data).  A  summary  of  statistics  for  this  data  is  found  in  Table  1 . 
Offsets  ranged  from  a  low  of  0  meters  (perfect  correspondence  between  field  &  ITD 
coordinates)  to  a  high  of  339  meters.  The  average  offset  for  all  observations  was  25 
meters.  When  the  GPS  measurement  error  (Q  value)  of  10  meters  is  included,  an 
average  offset  of  15  to  35  meters  is  established. 


Table  1.— Summary  of  ITD  Feature  Coordinate  Offsets 


U  obs. 
(E&N) 

OFFSET  (m) 
Low  Mean  High 

Std 
Dev.  (m) 

W%  offset 
50m 

W%  offset 
1 00m 

Q 

842 

0  25  339 

31 

73/8.67 

18/2.14 

10 

Q  -  measure  of  GPS  error  (circular)  in  meters 
E&N  -  Eastings  &  Northings 

Of  the  842  total  observations,  769  features  (91.3  %)  were  within  50  meters  of 
their  expected  locations.  This  is  within  the  expected  accuracy  of  the  1 : 50,000  Class 
B  TLM  (50  meters  circular  @  90%  confidence),  which  is  the  only  related  product 
with  a  formal  accuracy  statement.  Only  73  features  (8.7%)  had  field/ITD  offsets 
above  the  50-meter  TLM  baseline.  Further  examination  revealed  that  18  (2. 1%)  of 
the  73  had  offsets  over  100  meters.  As  expected,  93  percent  (68  of  73)  of  the  offsets 
over  50  meters  occurred  in  cartographically  controlled  data.  These  errors  were 
especially  prevalent  in  the  SD  (e.g.  dams,  canals),  OBS,  and  VEG  coverages  where 
55  of  the  73  offsets  were  observed.  Potential  rationales  for  the  higher  offsets  in  the 
carto-controlled  SD,  OBS,  and  VEG  coverages  (vs.  photo-controlled  ITD)  lie  in 
DMAs  collection/symbolization  techniques  for  this  data  as  described  below. 

In  carto-controlled  ITD,  horizontal  positioning  is  keyed  to  a  base  map.  Although 
many  features  are  collected  from  imagery,  they  are  still  subject  to  a  potential  offset 
of  50  meters  (map  accuracy).  Moreover,  SD  features  such  as  dams  and  canals, 
vegetation  polygons,  and  especially  obstacles  are  often  displaced  for  symbolization 
purposes.  Obstacles  such  as  embankments,  escarpments,  ditches  or  road  cuts  may 
be  substantially  displaced  to  ensure  the  correct  alignment  of  a  stream  or  road  bed. 
Moreover,  it  is  often  difficult  to  determine  the  exact  start  or  end  point  of  an  obstacle 
in  the  field. 

Vegetation  features  are  inherently  characterized  by  their  natural  variability. 
Furthermore,  ITD  vegetation  has  a  minimum  areal  size  requirement  resulting  in  the 
inclusion  of  many  different  vegetation  types  within  a  single  polygon.  The  resulting 
boundaries  can  vary  in  these  transition  areas  causing  offsets. 


305 


The  production  procedure  for  delineation  and  depiction  of  dams  and  canal 
end/break  points  (above/below  ground  transition)  is  unlike  that  of  other  SD  features 
in  which  positional  accuracy  is  strictly  upheld.  Dam  and  canal  lengths  are  often 
variable  since  they  are  roughly  plotted  to  scale  to  provide  a  relative  size.  More 
emphasis  is  placed  on  the  location/axis  of  these  features  rather  than  their  size/length. 

Figure  1  depicts  the  data  offset  for  an  actual  point  bridge  as  observed  on  the 
Gatesville,  TX  cell,  map  sheet  6446-1.  The  ITD  bridge  feature  is  approximately  25 
meters  from  the  road/stream  crossing  point  where  the  field  measurement  took  place. 
However,  the  actual  position  of  the  bridge  may  be  as  close  as  15  meters  or  as  far  as 
35  meters  from  its  intended  location  depending  on  the  GPS  error  introduced  in  the 
field. 

Table  2  illustrates  various  comparisons  between  the  carto-and  photo-controlled 
ITD  at  the  thematic  level.  Cartographic  feature  offsets  are  acceptable,  with  some 
coverages  performing  better  than  others.  For  example,  offsets  for  the  carto 
transportation  features  averaged  approximately  24  meters,  while  obstacles  offsets 
approached  70  meters.  The  mean  offset  for  the  carto  vegetation  features  (~23m) 
was  comparable  to  transportation,  but  exhibited  a  much  higher  95  percent  confidence 
interval  (-22.5  vs.  «4.5m).  This  indicates  increased  variability  for  the  vegetation 
features  and  a  potentially  lower  correlation  with  subsequent  field  observations. 

Overall  confidence  intervals  for  the  carto-controlled  features  ranged  from  ~4m 
(transportation)  to  -53  meters  (obstacles).  The  mean  offset  for  the  cartographic 
data  is  28.8  meters.  Conversely,  the  mean  offsets  of  the  photo-controlled  features 
were  very  low  across  all  thematic  layers.  They  ranged  from  a  low  of  10-12  meters 
(vegetation,  surface  drainage,  and  transportation  features)  to  a  high  of  21  meters 
(obstacles).  The  95  percent  confidence  intervals  are  fairly  consistent  across  features 
ranging  from  ~6  meters  (transportation)  to  »22  meters  (obstacles).  The  mean  offset 
for  the  photo  data  is  13.7  meters.  This  is  directly  attributable  to  the  increased  fidelity 
provided  by  a  solely  photogrammetric  production  technique. 

It  is  recognized  that  statistics/comparisons  for  several  of  the  feature  groupings  in 
Table  2  are  based  on  small  sample  sizes  (<30  observations).  However,  in  each  case 
the  t-statistic  exceeds  the  critical  value,  indicating  that  random  error  cannot  explain 
the  variance  observed  between  the  field  data  coordinates  and  those  of  ITD. 
Therefore,  there  is  a  significant  difference  between  the  populations  at  the  95  percent 
confidence  level. 

It  should  also  be  noted  that  all  of  the  ITD  data  used  for  this  study  was  collected 
over  areas  of  the  continental  United  States  (CONUS)  with  relatively  reliable  source 
materials.  Assuming  the  availability  of  comparable  sources,  findings  in  this  report 
should  be  extensible  to  any  location.  However,  it  is  unknown  to  what  extent,  if  any, 
foreign  or  other  ancillary  source  materials  may  affect  the  results  herein. 


306 


Table  2.— Carto/Photo  UD  vs.  Field  Data  Offsets  by  Feature  &  Associated  Statistics 


Photo-Controlled  Data  in  meters 


Measurement 
Type 

Average 
(Mean) 

Std. 
Dev. 

95%  Confidence 
int.  UJ  mean 

T 
Stat. 

CriL 
VaL 

Sample 
Size 

All  features 

1  J.  1  L 

1  Q  AQ 

1  C\  Qft     1  A  ^17 

Q  ft 
V.o 

1  A^ 
1  ,Oj 

1  OA 

Trans- 
portation 
features 

1 1  /CO 

n  to 

O  AO     1  ^  AO 

ft  17 
O.J  / 

1 .OO 

Obstacle 
features 

21.31 

30.85 

10.19-32.44 

3.91 

1.70 

32 

Surf.  Drainage 
features 

11.00 

11.57 

5.25-16.75 

4.04 

1.74 

18 

Vegetation 
features 

10.00 

9.56 

4.91  -  15.09 

4.19 

1.75 

16 

Carto-Controlled  Data  in  meters 


Measurement 
Type 

Average 
(Mean) 

Std. 
Dev. 

95%  Confidence 
Int  Of  Mean 

T. 
Stat 

CriL 
VaL 

Sample 
Size 

All  features 

28.76 

32.25 

26.27-31.25 

22.67 

1.65 

646 

Trans- 
portation 
features 

24.43 

24.97 

22.20  -  26.65 

21.61 

1.65 

488 

Obstacle 
features 

69.68 

75.57 

43.30  -  96.05 

5.38 

1.69 

34 

Surf.  Drainage 
features 

37.04 

30.05 

31.14-42.94 

12.45 

1.66 

102 

Vegetation 
features 

23.27 

25.42 

12.00-  34.54 

4.29 

1.72 

22 

307 


CONCLUSIONS 


1.  The  offset  of  well-defined  ITD  features  taken  as  a  whole  was  approximately  25 
meters.  Approximately  90  percent  of  these  features  were  within  50  meters  of  their 
expected  locations.  This  meets  the  accuracy  specification  of  50  meters  for  a  class 
B,  l:50,000-scale  TLM,  a  traditional  source  for  targeting  information. 

2.  Overall,  carto-controlled  ITD  feature  offsets  (29  meters)  were  higher  than  photo- 
controlled  ITD  feature  offsets  (14  meters).  Our  expectations  were  that  the  photo- 
controlled  ITD  would  perform  very  well,  as  it  did;  however,  performance  of  the 
carto-controlled  ITD  was  also  suprisingly  good. 

3.  The  positional  accuracy  of  ITD  features  was  determined  to  be  compatible  with 
the  positional  accuracy  of  hand-held  precise  GPS  equipment  currently  used  in  the 
field.  This  accuracy  compatibility  will  lessen  the  chance  for  confusion  and 
correlation  problems  such  as  could  occur  when  features  are  located  close  to  one 
another  on  the  ground  and  GPS-derived  locations  do  not  match  data  base  locations. 
This  problem  may  be  more  of  a  concern  with  lower  accuracy  data  sets  such  as 
VMAP  Level  1  (1:250,000  scale)  and  DCW  (1:1,000,000  scale). 

ACKNOWLEDGMENTS 

The  authors  wish  to  extend  special  appreciation  to  Mr.  John  Nedza,  Geographic 
Information  Lab,  Topographic  Engineering  Center  (TEC)  for  the  many  hours  of 
effort  he  provided  during  import  of  the  raw  Interim  Terrain  Data  (ITD)  data  and 
subsequent  conversion  to  Arc-Info  coverages.  His  outstanding  technical  expertise 
greatly  facilitated  data  analyses  and  played  a  critical  role  in  the  successful  completion 
of  this  study. 

The  authors  also  wish  to  thank  Mr.  Bill  Ryder,  Special  Studies  Division,  Digital 
Concepts  and  Anaysis  Center,  TEC  for  his  excellent  terrain  analysis  support  during 
the  field  collection;  Professor  Richard  Scott,  Rowan  College,  Glassboro,  N.J.  for 
advice  and  insight  about  the  statistical  processes  used  in  the  report;  and  Messrs. 
Robert  Gormley  and  David  Wu,  Special  Studies  Division,  Digital  Concepts  and 
Analysis  Center,  TEC  for  their  work  in  regard  to  research  and  compilation  of  the 
statistical  analyses. 


308 


Acquisition  of  Spatial  Data  by  Forest 
Management  Agencies:  a  Review 

Michael  J.C.  Weir1 

Abstract.  —  Forest  management  agencies  have  a  long  tradition  of 
surveying  and  mapping  for  the  acquisition  of  a  wide  range  of  spatial  data. 
These  include:  i)  external  (cadastral)  boundaries;  ii)  permanent,  internal, 
forest  boundaries;  iii)  forest  stand  and  other  impermanent  boundaries;  and 
iv)  roads  and  other  topographic  features.  In  recent  years,  the 
well-established  processes  of  graphical  map  compilation  and  revision  have 
been  increasingly  replaced  by  digital  mapping  techniques. 

Based  on  a  limited  survey,  together  with  published  information,  this  paper 
reviews  the  surveying  and  mapping  techniques  currently  employed  by 
forest  mapping  agencies  for  the  task  of  digital  database  establishment  and 
maintenance.  Accuracy  tolerances  specified  for  different  categories  of 
information  vary  considerably.  The  survey  showed  that  not  all  agencies 
apply  checks  to  ensure  that  specified  tolerances  are  achieved.  It  appears 
also  that  some  agencies  apply  techniques  whose  achievable  accuracy  is 
inconsistent  with  specified  tolerances. 


INTRODUCTION 

The  acquisition  of  spatial  data  to  assess  the  state  of  forest  resources  has  its 
origins  in  central  Europe  during  the  15th  century  (Griess,  1989).  Since  the  early 
19th  century,  when  forest  management  based  on  reliable  area  statistics  gradually 
became  common  practice  in  Europe,  forestry  agencies  have  employed  personnel 
and  established  facilities  for  the  specific  tasks  of  surveying  and  map  production. 
In  Canada  alone,  about  30  million  hectares  are  inventoried  annually  by  provincial 
forestry  agencies.  These  agencies  maintain  more  than  37,000  map  sheets  at  scales 
varying  from  1:10,000  to  1:20,000  (Leckie  and  Gillis  1995).  Spatial  data 
acquisition  by  forest  management  agencies  worldwide  therefore  accounts  for  a 
significant  portion  of  mankind's  total  effort  and  investment  in  resource  mapping. 

The  aim  of  this  paper  is  to  present  a  review  of  the  surveying  and  mapping 
techniques  currently  employed  by  forest  mapping  agencies  for  digital  database 
establishment  and  maintenance.  The  presented  information  is  mainly  based  on  data 


lecturer,  International  Institute  for  Aerospace  Survey  and  Earth  Sciences  (ITC),  Enschede.  The  Netherlands. 


309 


from  a  small  survey  of  forestry  agencies  known  to  use  GIS  in  production  (research 
organizations,  were  not  included).  The  aim  of  this  survey  was  to  determine:  i)  how 
forest  management  agencies  acquire  spatial  data  for  input  to  GIS;  ii)  the  accuracy 
tolerances  (if  any)  applied;  and  iii)  policies  concerning  the  updating  of  the  spatial 
database.  A  questionnaire  covering  these  topics  was  sent  to  33  agencies  in  various 
parts  of  the  world  and  25  replies  were  received. 

The  remainder  of  the  paper  is  divided  into  three  sections.  The  first  section 
briefly  describes  the  main  categories  of  spatial  data  required  by  forest  management 
agencies.  The  technology  used  by  the  25  surveyed  agencies  to  acquire  these  data 
is  summarized.  In  the  second  section,  specified  accuracy  tolerances  are  examined 
in  the  light  of  the  surveying  and  mapping  technology  employed.  The  paper 
concludes  with  some  remarks  on  spatial  data  accuracy  and  error  modelling  in 
forest  management. 

SPATIAL  DATA  ACQUISITION 

Although  maps  are  probably  the  documents  most  frequently  consulted  during 
forest  operations  (Moser  1970),  technical  details  their  production  are  not  widely 
published,  neither  in  the  forestry  nor  in  the  surveying  and  mapping  literature. 
Leckie  and  Gillis,  (op.  cit.)  provide  extensive  details  of  current  forest  mapping 
practice  in  Canada.  A  recent  report  (Kennedy  et  al.  1994)  summarizes  forest 
mapping  activities  in  Europe.  Probsting  (1994)  provides  a  useful  review  of 
applications  of  aerial  photography  in  European  forest  mapping. 

Information  content 

The  management  of  forests  and  forest  land  requires  a  wide  range  of  spatial 
information,  usually  obtained  from  a  variety  of  sources.  For  example,  a  GIS  to 
support  the  formulation  of  "Indicative  Forestry  Strategies"  in  Scotland  uses  34 
sources  of  biophysical  and  administrative  data  (Aspinall  et  al.  (1993).  Although 
requirements  vary  from  agency  to  agency  depending  on  forest  conditions  and 
management  goals,  maps  and  GIS  databases  generally  comprise  four  main 
categories  of  spatial  data,  namely:  cadastral  boundaries,  "internal"  administrative 
(for  example,  compartment)  boundaries,  topographic  details  and  information  on 
the  forest  resources.  Information  on  spatial  data  acquisition  by  the  25  management 
agencies  was  compiled  for  each  of  these  categories  of  information  (table  1).  It 
should  be  noted  that  the  data  in  the  table  refer  only  to  the  number  of  instances 
where  a  particular  technique  is  mentioned  as  being  used.  The  data  do  not  consider 
the  area  concerned  or  the  amount  of  equipment  employed.  Furthermore,  many 
agencies  apply  more  than  one  technique  to  acquire  certain  categories  of  spatial 
data. 


310 


Table  I.  -  Sources  of  spatial  data  required  by  25  forest  management  agencies 

(number  of  occasions  mentioned). 


Source 

Cadastral 

Topographic 

Internal 

Forest 

boundaries 

features 

boundaries 

boundaries 

Supplied  by  other 

agencies  as: 

maps 

12 

15 

5 

2 

digital  data 

8 

6 

1 

1 

Acquired  by  the  forest 

agency  using: 

Land  survey 

7 

6 

9 

9 

GPS 

1 

5 

1 

4 

Graphical  transfer 

1 

6 

6 

13 

from  aerial  photographs 

Graphical  stereoplotting 

1 

2 

4 

6 

Digital  stereoplotting 

1 

5 

4 

7 

Orthophotography 

3 

7 

4 

6 

Satellite  imagery 

2 

1 

2 

Not  required 

1 

6 

Cadastral  information 

An  exact  knowledge  of  the  area  and  limits  of  the  forest  land  is  essential  for 
proper  management.  In  many  parts  of  Europe,  large  scale  cadastral  maps  are  the 
source  documents  for  the  compilation  of  the  "basic  forest  map"  (German 
Forstgrundkarte).  Forest  management  agencies  generally  obtain  information  on 
land  ownership  boundaries  from  cadastral  survey  agencies.  In  some  countries, 
however,  the  forest  management  agency  itself  may  have  official  cadastral  survey 
tasks  such  as  the  determination  and  demarcation  of  parcel  boundaries  on  forest 
land  (Wander  1981). 

Topographic  information 

General  topographic  information  is  required  for  planning  forest  operations  such 
as  road  construction  and  timber  harvesting  and  also  to  provide  a  metric  framework 
within  which  sample  plots  for  forest  inventory  can  be  located.  Most  importantly, 
details  such  as  roads  or  drainage  features  may  define  (parts  of)  forest  stand 
boundaries.  Topographic  information  therefore  forms  an  integral  part  of  any 
(graphical  or  digital)  forestry  database  rather  than  simply  a  background  against 
which  details,  such  as  forest  types,  are  mapped.  The  survey  of  25  forest  agencies 
showed  that  they  obtain  topographic  information  primarily  from  other  agencies. 
However,  surveys  by  forestry  agencies  themselves  are  generally  required  to  map 
additional  details  not  shown  on  general  purpose  topographic  maps  and  "new" 
features  such  as  recently  constructed  forest  roads.  Land  surveying,  GPS  and  a 


311 


range  of  photogrammetric  methods  are  all  employed  for  this  purpose.  When 
relatively  approximate  methods  of  graphical  transfer  from  aerial  photographs  are 
employed,  these  additional  topographic  features  are  likely  to  be  acquired  with  a 
spatial  accuracy  lower  than  that  achieved  by  the  national  mapping  agency,  leading 
to  inconsistencies  in  spatial  data  quality  among  different  sources. 

Permanent  boundaries  established  by  forest  agencies 

Many  forestry  organizations  establish  and  maintain  a  "quasi  cadastral"  survey 
of  the  land  under  their  jurisdiction.  This  may  take  the  form  of  compartments  or 
timber  concession  areas,  the  boundaries  of  which  must  be  demarcated,  surveyed 
and  mapped.  The  survey  showed  that  a  wide  range  of  surveying  and  mapping 
techniques  are  employed  to  acquire  data  on  permanent  "internal"  forest  boundaries. 
Unlike  "external"  property  boundaries  which,  for  reasons  of  accuracy,  are 
primarily  surveyed  in  the  field,  permanent  forest  boundaries  need  not  be  surveyed 
to  cadastral  mapping  standards.  Photogrammetric  techniques  are  therefore  feasible 
and,  in  fact,  apparently  predominate  over  field  surveys. 

Forest  information 

Information  on  stand  boundaries  and  other  "specialized"  forest  information  is 
not  generally  available  from  existing  sources.  Acquisition  of  this  category  is 
therefore  one  of  the  main  tasks  of  any  forest  surveying  and  mapping  agency.  Just 
over  half  of  the  agencies  employ  simple  methods  of  graphical  transfer  from  aerial 
photographs  for  this  purpose.  The  operating  costs  of  these  "simple"  graphical 
methods  are  not  significantly  lower  than  those  associated  with  the  use  of  rigorous 
photogrammetric  methods  (Weir  1981).  It  is  therefore  surprising  to  find  this  labour 
intensive  approach  still  so  widely  used  by  forestry  agencies  in  developed 
countries.  This  technique  is  employed  by  many  provincial  forestry  agencies  in 
Canada  (Leckie  and  Gillis,  op.  cit.),  although  some  are  now  replacing  it  by  more 
efficient  photogrammetric  methods  such  as  digital  mono  plotting. 

As  table  1  indicates,  forest  management  agencies  make  only  limited  use  of 
satellite  imagery  for  forest  type  mapping.  Clearly,  the  spatial  resolution  of  the 
present  generation  of  earth  observation  satellites  is  inadequate  for  mapping  at  the 
scale,  level  of  detail  and  positional  accuracy  generally  required  by  agencies 
practicing  intensive  forest  management.  In  many  developing  countries,  however, 
satellite  imagery  may  be  the  only  source  of  timely  spatial  data  on  forest  resources, 
although  informal  discussion  with  foresters  from  these  countries  suggest  that  there 
too,  aerial  photography  remains  the  principal  source  of  spatial  data. 

Frequency  of  revision 

Unlike  a  soil  or  geological  map  which,  if  carefully  produced,  may  serve  as  a 
useful  document  for  many  decades,  a  forest  management  map  must  be  updated  at 


312 


regular  intervals.  Consequently,  map  revision  is  an  important  aspect  of  forest 
management  mapping.  In  the  intensively  managed  forest  of  western  Europe,  where 
management  plans  are  traditionally  prepared  every  ten  years,  management  maps 
are  revised  on  a  corresponding  10-year  cycle. 

Nowadays,  map  revision  is  increasingly  carried  out  as  a  process  of  (digital) 
spatial  database  maintenance.  Table  2  gives  details  of  the  frequency  of  updating 
employed  by  16  of  the  surveyed  agencies  and  indicates  a  tendency  towards 
continuous  or  annual  updating.  Corrie  et  al.  (1994)  describe  the  database 
maintenance  procedures  employed  by  a  timber  company  for  2.4  million  hectares 
of  productive  forest  in  British  Columbia.  The  database,  which  includes  250  base 
and  thematic  layers,  is  continuously  updated,  with  all  changes  distributed  to  local 
offices  on  a  nightly  basis. 


Table  2.  -  Frequency  of  database  updating  by  16  forest  management  agencies. 


Cadastral 

Topographic 

Internal 

Forest 

boundaries 

features 

boundaries 

boundaries 

Continuous 

7 

9 

9 

5 

Several  times  per  year 

1 

1 

Annually 

3 

3 

3 

4 

Every  2-5  years 

2 

1 

1 

1 

Every  10  years 

2 

1 

3 

5 

In  fact,  experience  with  database  updating  is  still  limited.  At  the  time  of  the 
survey  (mid- 1994),  only  two  systems  had  been  in  operation  for  more  than  ten 
years.  Eight  agencies  had  being  using  their  system  for  less  than  five  years  and 
some  agencies  had  not  yet  done  any  updating.  Interestingly,  a  few  agencies 
indicated  that  forest  stand  information  is  updated  by  means  of  a  complete 
remapping  rather  than  by  considering  only  the  actual  changes.  Although  this 
would  appear  to  negate  the  advantages  of  having  a  digital  database,  there  may  be 
a  limit  to  the  amount  of  revision  which  can  take  place.  One  agency,  which  has 
been  employing  GIS  since  1977,  updates  the  database  on  a  2-year  cycle  and 
makes  completely  new  overhaul  every  ten  years. 

ACCURACY  STANDARDS 

In  forest  management,  spatial  data  are  required  to  provide  not  only  locational 
information  but  also,  most  importantly,  area  data  for  the  calculation  of  timber 
volume  and  other  quantities.  Ideally,  therefore,  the  errors  associated  with  these 
area  data  should  be  known.  Leckie  and  Gillis  (op.  cit.  p81)  report  that  most 
Canadian  forest  inventory  maps  are  considered  to  have  accuracies  "in  the  10  to 
25m  range",  although  as  Gemmell  et  al.  (1991,  p221)  note,  "most  existing  forest 


313 


inventories  do  not  have  estimates  of  the  accuracies  of  their  polygon  labels  and 
boundaries". 

The  survey  of  25  forest  management  agencies  suggested  that,  among  the  four 
main  categories  of  information,  cadastral  boundaries  tend  to  have  relatively 
stringent  tolerances.  As  indicated  in  table  1,  however,  this  category  of  data  is 
generally  provided  by  other  agencies  according  to  predetermined  standards.  The 
other  three  categories  of  information  are  mostly  all  acquired  to  the  same  standards 
(although  some  agencies  relax  the  tolerances  for  -  relatively  uncertain  -  forest 
stand  boundaries).  Actual  values  vary  markedly  from  one  agency  to  another  or  are 
expressed  as  a  range  (for  example,  10  to  30m)  within  which  map  accuracy  should 
fall.  Tolerances  for  stand  boundaries  set  by  individual  forest  management  agencies 
in  Europe  are  generally  in  the  range  2  -  10m  rmse  relative  to  nearby  fixed  detail. 

Although  virtually  all  of  the  25  surveyed  agencies  indicated  that  they  set 
accuracy  standards  for  the  acquisition  of  spatial  data,  almost  half  (12)  indicated 
that  they  do  not  apply  any  quality  control  measures  to  check  if  accuracy  tolerances 
are  actually  being  met.  Interestingly,  eight  of  these  12  agencies  employ  simple 
graphical  methods  for  mapping  forest  boundaries  from  aerial  photographs.  Tests 
using  a  variety  of  instruments,  photo-scales  and  operators  (Weir,  op.  cit.)  have 
shown  that  the  accuracy  of  these  methods  is  highly  sensitive  to  terrain  conditions, 
base  map  quality  and  operator  skill.  Their  performance  (under  optimal  conditions, 
0.8  -  1.3mm  rms  plotting  error)  is  generally  inconsistent  with  required  standards. 

CONCLUSIONS 

This  paper  has  shown  that  many  forestry  agencies  already  work  to,  or  at  least 
aim  at,  "near  professional"  standards  of  surveying  and  mapping.  In  some  cases, 
however,  there  is  a  need  to  define  these  standards  more  clearly  and,  in  particular, 
to  apply  appropriate  quality  control  measures. 

In  forestry,  spatial  data  acquisition  is  not  simply  a  matter  of  map  production. 
Data  on  the  areas  of  individual  forest  stands  form  an  essential  input  for  forest 
management  planning.  Spatial  data  quality  therefore  has  a  direct  impact  on  forest 
management  operations  and  business  decisions  (Prisley  1994).  The  establishment 
and  maintenance  of  a  spatial  database  can  represent  a  considerable  investment  for 
a  forestry  agency,  and  it  may  be  difficult  to  justify  measures  to  improve  spatial 
data  quality  (Keefer  1994). 

In  recent  years,  GIS  researchers  have  laid  much  of  the  ground  work  needed 
to  handle  spatial  error  in  natural  resource  databases.  A  lot  of  the  research  on  error 
propagation  (see,  for  example,  Heuvelink  1993)  is,  however,  concerned  with 
situations  in  which  the  accuracy  of  the  basic  data  is  predetermined.  Because  they 


314 


acquire  a  large  part  of  their  primary  data  themselves,  forest  management  agencies 
are  in  the  fortunate  position  of  being  able  to  chain  backwards  through  the  various 
GIS  operations  in  order  to  define  the  accuracy  of  primary  data  acquisition  needed 
to  achieve  the  desired  quality  of  the  information  product.  In  this  way,  the  methods 
and  costs  of  spatial  data  acquisition  can  be  better  matched  to  the  value  of  the 
resource  being  mapped. 

ACKNOWLEDGEMENT 

This  paper  is  based  on  information  about  spatial  data  acquisition  kindly 
supplied  by  forestry  agencies  in  many  parts  of  the  world. 

REFERENCES 

Aspinall,  R.J.,  Miller,  D.R.  and  Birnie,  R.V.  1993.  Geographical  information 

systems  for  rural  land  use  planning.  Applied  Geography,  13:54-66. 
Corrie,  G.,  Reedijk,  W.  and  Lohia,  K.  1994.  A  GIS  to  improve  the  competitive 

advantage  of  TimberWest  Forest  Limited.  Proc.  GIS'94,  Vancouver:83-88. 
Gemmell,  F.,  Goodenough,  D.G.,  Fung,  K.  and  Kushigbor,  C.  1991.  Resource 

spatial  and  attribute  information  extraction  from  remotely  sensed  data.  Proc. 

Spatial  Data  2000,  Oxford :22 1-231. 
Griess,    O.    1989.    Die    Forsteinrichtung,    ein    klassisches  geographisches 

Informationssystem.  Proc.  Symposium  on  New  Technologies  in  Forest 

Management  Planning,  IUFRO  S4.04,  Luxemburg :8pp. 
Heuvelink,  G.B.M.  1993  Error  Propagation  in  Quantitative  Spatial  Modelling: 

Applications  in  Geographical  Information  Systems.  Nederlandse  Geografische 

Studies  163,  KNAG,  Utrecht:  15  lpp. 
Keefer,  B.J.  1994.  Impact  of  spatial  accuracy  on  business  decisions:  a  forest 

management  perspective.  Proc.  International  Symposium  on  the  Spatial 

Accuracy  of  Natural  Resource  Data  Bases,  Williamsburg:  18-23. 
Kennedy,  P.J.,  Paivinen,  and  Roihuvuo,  L.  (eds.)  1994.  Proc.  International 

Workshop  on  Designing  a  System  of  Nomenclature  for  European  Forest 

Mapping.  European  Commission,  Luxemburg:455pp. 
Leckie,  D.G.  and  Gillis,  M.D.  1995.  Forest  inventory  in  Canada  with  emphasis  on 

map  production.  Forestry  Chronicle,  71(l):74-88. 
Moser,  W.  1970.  Wie  genau  muB  eine  Forstkarte  sein?  Allgemeine  Forstzeitung, 

81(8):205-207. 

Prisley,  S.  P.  1994.  Why  natural  resource  information  managers  must  be 
concerned  about  spatial  accuracy.  Proc.  International  Symposium  on  the 
Spatial  Accuracy  of  Natural  Resource  Data  Bases,  Williamsburg:24-34. 

Probsting,  T.  1994.  Einsatz  von  Luftbildern  in  der  Forstwirtschaft  in  Europa  -  Ein 
Uberblick.  In  T.  Probsting  (ed.)  Photogrammetrie  &  Forst.  Albert-Ludwigs- 
Universitat,  Freiburg:277-296. 


315 


Wander,  R.   1981.  Die  Forstvermessung  in  Landesteil  Baden.  Allgemeine 

Vermessungsnachrichten  8/9:329-338. 
Weir,  M.J.C.  1981.  Simple  plotting  instruments  for  resource  mapping.  Proc. 

International  Conference  on  Matching  Remote  Sensing  Technologies  and  their 

Applications,  London:223-232. 

BIOGRAPHICAL  SKETCH 

Michael  Weir  is  a  lecturer  at  the  International  Institute  for  Aerospace  Survey 
and  Earth  Sciences  (ITC),  Enschede,  The  Netherlands.  He  is  mainly  responsible 
for  teaching  surveying,  photogrammetry  and  GIS  to  foresters  and  other  resource 
specialists  from  developing  countries. 


316 


Attribute  and  Positional  Accuracy 
Assessment  of  the 
Murray  Darling  Basin  Project, 

Australia 

Fitzgerald,  R.W.1,  Ritman,  K.T.  &  Lewis,  A. 


Abstract:  The  Murray  Darling  Basin  comprises  a  land  area  of 
1,058,000  km2  covering  a  substantial  portion  of  SE  Australia  and 
encompassing  Australia's  largest  river  system.  A  basin  wide  woody 
vegetation  dataset  has  been  assembled  from  LANDSAT  TM  imagery 
supplemented  with  aerial  photography  at  a  nominal  scale  of  1 : 100,000. 

The  attribute  accuracy  assessment  method  is  based  on  a  multi-stage 
systematic  sample  design.  A  rectangular  grid  is  placed  over  the  entire 
Murray  Darling  Basin  dataset  and  at  each  primary  sample  point,  a 
secondary  grid  is  created  formed  from  a  square  contiguous  set  of  aerial 
photos. 

A  half  tone,  grey  scale  transparency  covering  approximately  10km2 
is  generated  from  satellite  digital  imagery  (LANDSAT  TM)  for  each 
primary  grid  point  at  a  contact  scale  matching  the  available  air  photos. 
Overprinted  on  this  transparency  are  the  secondary  grid  points  as  patch 
size  sampling  frames.  The  air  photos  are  then  visually  co-registered 
underneath  the  transparency  with  Landcover  features  from  the 
LANDSAT  TM  image.  Attribute  data  (presence  or  absence  of  woody 
vegetation)  is  collected  directly  from  the  air  photos  for  each  secondary 
grid  point  at  4  different  patch  sizes. 

Positional  accuracy  is  assessed  by  recording  40  or  more  ground 
control  points  from  the  1:100,000  map  sheet  containing  the  primary 
grid  point.  These  Eastings  and  Northings  are  compared  to  their  position 
in  the  LANDSAT  TM  image. 

The  outcome  of  the  accuracy  assessment  is  a  Basin  wide  attribute 
and  positional  accuracy  statement  and  spatial  variability  contour  maps 
of  attribute  and  positional  accuracy  as  GIS  overlays. 

INTRODUCTION 

This  paper  discusses  the  attribute  and  positional  accuracy  assessment 


Infoplus.  Po  Box  125.  Queanbeyan.  NSW  2620.  Australia.  Fax:  +61  (0)6  299  1331.  E-mail 
bfitzgerQipcug.  org.  au 


317 


methodology  developed  for  the  Murray  Darling  Basin  Project2  (MDBP).  The  Murray 
Darling  Basin  (MDB)  covers  an  area  exceeding  440  x  1 : 1 00,000  map  sheets  with  a 
total  area  of  1,058,000  km2.  The  diverse  range  of  vegetation  and  land  forms  creates 
challenges  in  terms  of  methodology  and  logistics. 

The  initial  focus  of  the  accuracy  assessment  method  include  three  stratification 
hypothesise.  The  first  is  that  the  local  geographic  variations  in  the  classification 
accuracy  is  a  function  of  vegetation  type,  terrain  and  substrate.  Second,  the  smaller 
the  patch  of  woody  vegetation,  the  lower  the  attribute  accuracy.  Thirdly,  the  overall 
accuracy  percentage  is  proportional  to  the  percentage  of  vegetation  cover.  If  the 
vegetation  cover  is  patchy,  then  the  ratio  between  the  length  of  a  boundary  around 
a  patch  polygon  and  it's  enclosed  area  is  of  interest.  This  ratio  was  examined  by 
Crapper  et.al.  (1986). 

The  positional  and  accuracy  assessment  methodology  will  be: 

a.  Statistically  sound,  practical,  inexpensive  to  implement,  easily 
understood  by  project  staff  and  portable  to  MDBP's  GIS; 

b.  Produce  scalable  attribute  and  positional  accuracy  assessments  of  the 
woody/  non  woody  vegetation  dataset  as; 

i.  Attribute  accuracy  assessment  statistics  (error  matrix,  user 
and  producer  accuracies  and  Kappa  statistic)  at  different 
spatial  scales; 

ii.  Spatial  variability  maps  of  attribute  and  positional  accuracy 
for  the  entire  MDBP. 

A  brief  examination  of  the  MDB  woody/non-woody  GIS  layer  provided  an 
insight  into  data  quality  and  processing  standards.  The  details  of  these  standards 
including  lineage  are  outlined  in  the  Draft  specification  (Ritman,  1995). 

The  Victorian  &  South  Australian  groups  have  examined  attribute  accuracy 
assessment  methodologies.  The  work  of  Czaplewski  et.al.  (1992)  proved  to  be  the 
most  substantial  and  well  documented  attribute  accuracy  assessment  methodology 
available.  They  used  aerial  photo  interpretation  as  their  pseudo  ground  truth  and  a 
systematic  sample.  Tadrowski  et.al.  (1990)  in  South  Australia  investigated 
supervised  classification  aided  by  manual  photo  interpretation  and  a  stratified  simple 
random  sample  for  attribute  accuracy  assessment. 

The  production  of  the  digital  woody  vegetation  dataset  at  a  nominal  Basin- wide 
scale  of  1:100,000  was  achieved  by  a  two  stage  process  (Ritman,  1995).  Woody 
vegetation  is  defined  as  any  perennial  vegetation  having  a  height  exceeding  2m  and 
a  density  greater  than  20%  crown  cover  (McDonald  et.al.,  1990). 

The  first  stage  was  the  digital  classification  from  LANDSAT  TM  imagery  of  a 
template  of  only  woody  vegetation.  This  method  is  based  on  that  of  Gilbee  & 
Goodson  (1992)  and  comprises  an  unsupervised  100  ISO  class  cluster  analysis  of 


This  paper  is  a  product  of  a  consulting  project  titled  "Attribute  Accuracy  Assessment  for  Project  M305'\ 
DLWC,  September  1995,  Murray  Darling  Basin  Project  M305. 


318 


LANDSAT  TM  imagery,  followed  by  manual  aggregation  based  on  field  input, 
aerial  photos  and  ancillary  data.  In  addition  a  filter  (an  ARCINFO  AML  process) 
designed  to  remove  patches  of  cells  diagonally  and  orthogonally  connected,  created 
by  Dr.  K.  Ritman,  was  passed  over  the  resulting  woody  vegetation  layer  to  remove 
unconnected  vegetation  patches  less  than  0.25  ha..  The  resultant  woody  dataset  is  in 
raster  format  with  25  x  25m  pixels. 

The  next  stage  was  either  a  manual  or  digital  classification  of  vegetation  structural 
elements  such  as  genus,  density  class  and  growth  form.  Only  the  woody  vegetation 
layer  is  subject  to  the  attribute  and  positional  accuracy  methodology  described  in  this 
paper. 

A  BRIEF  REVIEW  OF  THE  LITERATURE  AND  PREVIOUS 

STUDIES. 

Attribute  classification  accuracy  is  usually  assessed  by  constructing  a  contingency 
table  of  a  classified  map  versus  ground  truth  or  reference  data  (Congalton,  1991; 
Veregin  ,1989).  The  resulting  error  or  confusion  matrix  C  is  a  k  x  k  matrix  where 
k  is  the  number  of  discreet  classes  in  the  classification  scheme.  In  the  case  of  the 
MDB  woody/non- woody  dataset,  k  =  2. 

The  most  commonly  used  index  of  attribute  classification  accuracy  is  the  overall 
accuracy  percent  (OA%).  Confidence  limits  for  the  OA%  can  be  constructed  easily 
from  either  the  binomial  distribution  or  the  normal  approximation  to  the  binomial 
distribution.  Van  Genderen  et.al.  (1978)  logically  extends  the  use  of  the  binomial 
confidence  limits  to  estimate  sample  sizes  given  expected  classification  accuracy  and 
confidence  levels. 

The  OA%  is  a  simple  index  of  classification  accuracy  which  has  its  limitations. 
The  OA%  can't  differentiate  between  errors  of  omission  and  commission.  Also  it 
can  not  reliably  be  used  to  compare  the  performance  of  different  error  matrices  with 
different  sample  sizes  as  well  as  not  being  able  to  account  for  correct  classification 
by  chance  alone. 

One  of  the  best  methods  developed  to  overcome  these  limitations  is  the  Kappa 
statistic  discussed  extensively  by  Congalton  (1991)  and  Fitzgerald  &  Lees  (1994a). 
It  statistically  quantifies  the  level  of  agreement  and  has  been  shown  to  give  a  less 
biased  estimate  of  classification  accuracy  than  the  OA%  (Rosenfield  &  Fitzpatrick- 
Lins,  1986). 

The  effects  of  sampling  schemes  on  classification  accuracies  especially  when 
viewed  across  the  spectral,  spatial,  environmental,  taxonomical  and  temporal 
domains  can  induce  substantial  bias  into  classification.  Congalton  (1988,  1991) 
compares  the  relative  effects  of  five  sampling  schemes  including  random  and 
stratified  random  sampling  on  classification  accuracy.  Franklin  &  Hiernaux  (1991) 
discuss  the  effect  of  sampling  schemes  on  woody  vegetation  classification  while 
Fitzgerald  &  Lees  (1994b)  discuss  scale  and  its  relationship  to  floristic  structure. 


319 


High  accuracy  surveying  standards  exist  for  assessing  the  accuracy  of  topographic 
maps  in  all  three  dimensions.  The  state  and  national  mapping  agencies  are 
responsible  for  the  surveying  standards  and  integrity  of  the  map  base.  The  implicit 
assumption  made  in  this  study  is  that  this  map  base  is  generally  correct3. 

Positional  accuracy  quantifies  the  accuracy  of  feature  locations  after  various 
image  processing  and  GIS  transformations  have  been  applied.  A  number  of  tests  are 
available  to  assess  positional  accuracy  including  deductive  estimates,  internal 
evidence  checks,  comparisons  to  source  documents  and  reference  to  independent 
sources  of  higher  accuracy  (Veregin,  1989).  The  latter  is  the  most  desirable  and  the 
one  used  in  this  study.  The  independent  source  is  the  AUSLIG  1 : 100,000  map  base 
series. 

Spatial  variability  maps  of  both  attribute  and  positional  accuracy  will  be  produced 
for  the  MDB  from  this  accuracy  assessment  methodology.  The  12  x  18  primary 
systematic  grid  (described  below)  will  contain  the  derived  data  values  of  OA%  and 
Kappa  (attribute  accuracy)  and  2D-RMS  and  CEP  (positional  accuracy).  These 
gridded  values  will  then  be  interpolated  to  a  surface  from  which  a  contour  map 
(isometric  lines)  of  attribute  and  positional  accuracy  will  be  produced. 

The  accuracy  of  this  interpolation  is  dependant  on  the  number  and  spatial 
distribution  of  the  observed  sample  values.  Systematic  sampling  (aligned  or 
unaligned)  is  the  best  of  all  the  sampling  techniques  tested  to  minimise  the  effects 
of  spatial  distribution  on  contouring  (Veregin,  1989). 

SAMPLE  DESIGN 

The  constraints  on  the  sample  design  for  the  attribute  and  positional  accuracy 
assessment  for  the  MDBP  were  as  follows: 

a.  The  extent  of  the  MDB  (1 ,058,000km2)  precludes  field  checking  as 
the  major  source  of  ground  truth.  Aerial  photo  interpretation  is  the 
only  practical  means  in  this  case; 

b.  The  design  must  be  practical,  simple  and  expedient  to  implement. 
Specifically  the  handling  of  maps,  air  photos,  air  photo  run  maps 
along  with  satellite  imagery  should  be  handled  as  efficiently  as 
possible  with  the  minimum  number  of  each  being  accessed  as 
possible; 

c.  Attribute  accuracy  must  be  assessed  at  a  minimum  patch  size  of  1 
hectare  across  the  entire  basin.  The  patch  size  scale  effect  on 
classification  accuracy  should  be  assessed  if  possible  at  scales  of 
0.25,  1.00,  4.00  and  9.00  ha.; 


This  assumption  proved  incorrect.  Field  experience  of  project  staff  demonstrated  that  the  map  base  is  not 
always  reliable.  Combined  with  budgetary  restrictions,  the  positional  accuracy  component  of  the 
methodology  was  cancelled  in  the  implementation  phase. 


320 


d.        The  results  of  the  accuracy  assessment  must  be  statistically 
defensible  and  easily  interpreted  by  the  end  user  community. 
With  these  constraints  in  mind  a  review  of  the  literature  suggested  that  the  four 
best  contenders  for  the  sample  design  were:  Simple  random  sampling  (SRS); 
Stratified  simple  random  sampling  (STSRS);  Cluster  sampling;  Systematic  sampling 
(SYS). 

Simple  random  sampling  has  the  advantage  of  being  the  easiest  to  construct. 
Implementation  especially  in  the  field  can  be  problematic.  Statistically,  the  estimates 
produced  are  easily  produced  and  are  robust.  For  the  purposes  of  the  MDBP,  SRS 
was  considered  impractical  to  implement  on  such  a  large  scale  dataset. 

Stratified  SRS  is  the  most  often  used  design  as  judged  from  literature.  It  is 
statistically  more  efficient  than  SRS  (Cochran,  1977),  can  produce  less  biased  results 
than  SRS  and  is  a  little  easier  to  implement  (Congalton,  1988:  Janssen  &  van  der 
Waal,  1994:  Van  Genderen  et.al.,  1978).  Constructing  a  stratified  SRS  is  more 
complex  than  a  SRS  by  itself  and  stills  suffers  from  the  problems  of  access  to  ground 
truth  sites.  The  claimed  statistical  gains  over  SRS  are  also  very  dependent  on  the 
spatial  autocorrelation  structure  of  the  dataset  (Congalton,  1988)  which  is  unknown. 
Stratified  SRS  was  not  considered  any  more  practical  to  implement  than  SRS  for  the 
MDBP. 

Cluster  sampling  is  the  preferred  design  when  cost  of  access  to  the  ground  truth 
sites  is  at  a  premium.  Moisen  et.al.  (1994)  showed  that  cluster  sampling  with  a  fixed 
cost  had  a  higher  relative  efficiency  than  either  systematic  or  simple  random 
sampling.  The  construction  and  implementation  of  cluster  sampling  are  more 
complex  than  either  stratified  or  simple  random  sampling.  Once  again  the 
practicalities  of  this  design  precluded  its  use  in  the  MDBP. 

Systematic  samples  have  as  their  most  attractive  feature  their  ease  of  construction 
and  implementation.  Systematic  sampling  is  particularly  suited  to  spatial  problems 
involving  two  dimensions  or  more  as  noted  by  Cochran  (1977).  Cochran  (1977)  and 
Congalton  (1988)  propose  that  unaligned  systematic  sampling  is  superior  to  aligned 
or  centred  systematic  sampling. 

A  large  number  of  authors  and  studies  have  used  systematic  sampling.  Goodchild 
et.al.  (1991)  utilised  a  single  stage  systematic  sample  of  1347  sites  in  the  CALVEG 
study.  Czaplewski  et.al.  (1992)  used  a  systematic  sample  of  363  plots  for  the 
Victorian  CNR  Tree  cover  project. 

The  attribute  accuracy  assessment  method  recommended  for  the  MDBP  is  a  two 
stage  systematic  sample  with  subsampling  units  of  equal  size.  The  sampling  unit  is 
a  woody  vegetation  patch.  The  sample  design  has  the  practical  advantage  of 
simplifying  the  acquisition  and  handling  of  the  aerial  photos  used  to  acquire  the 
pseudo  ground  truth  values  for  the  attribute  accuracy  assessment. 

In  the  first  sampling  stage,  a  12  x  18  rectangular  primary  sampling  grid  (216 
primary  grid  points)  is  placed  over  the  MDB  dataset.  This  grid  is  oriented  40°  from 
North  to  incorporate  linear  trends  in  Landcover  categories. 


321 


At  each  primary  sample  point,  a  7x7  secondary  grid  (49  secondary  grid  points) 
is  created  orientated  to  the  AMG  grid.  Thus  the  number  of  secondary  sample  points 
will  be:  12x18x7x7  =  10,584.  Due  to  the  irregular  outline  of  the  MDB, 
approximately  10%  of  the  primary  grid  point  will  fall  outside  of  the  Basin  and  will 
be  excluded.  Thus  the  final  sample  size  will  be  less  than  10,584,  possibly  around 
9,500.  The  sampling  unit  is  a  woody  vegetation  patch. 

This  two  stage  design  has  the  added  advantage  of  being  scalable.  The  grid  cell 
size  at  either  sample  stage  can  be  varied  to  reflect  desired  confidence  limits,  budget 
and  time  constraints4.  The  presence/  absence  of  woody  vegetation  at  the  4  patch  sizes 
(0.25,  1,  4  and  9  ha.)  will  be  collected  at  each  Secondary  grid  point.  This  data  will 
be  used  to  investigate  the  effect  of  patch  size  on  classification  accuracy. 

To  collect  the  attribute  information,  a  half  tone,  grey  scale  transparency  covering 
approximately  100km2  is  generated  from  satellite  digital  imagery  (LANDS AT  TM) 
for  each  primary  grid  point  at  a  contact  scale  matching  the  available  air  photos. 
Overprinted  on  this  transparency  are  the  secondary  grid  points  as  patch  size  sampling 
frames.  The  air  photos  are  then  visually  coregistered  underneath  the  transparency 
with  Landcover  features  from  the  LANDS  AT  TM  image  (MDBC,  1995). 

Attribute  data  (presence  or  absence  of  woody  vegetation)  is  collected  directly 
from  the  air  photos  for  each  secondary  grid  point  at  4  different  patch  sizes.  This 
information  forms  the  pseudo  ground  truth  values  for  comparison  to  the  LANDSAT 
TM  derived  vegetation  layer  (woody/  non  woody).  The  attribute  data  for  each  of  the 
49  secondary  grid  points  is  compiled  and  crosstabulated  with  the  classified 
LANDSAT  TM  woody  vegetation  values.  An  OA%,  user  and  producers  accuracy 
along  with  the  Kappa  statistic  are  derived  from  this  error  matrix  to  assess  the 
attribute  accuracy  at  each  primary  grid  point.  These  estimates  are  statistically  valid 
at  the  secondary  grid  point  level. 

Positional  accuracy  is  assessed  by  recording  40  or  more  ground  control  points 
(GCPs)  from  the  1:1 00000  map  sheet  containing  each  primary  grid  point.  These 
Eastings  and  Northings  are  compared  to  their  position  in  the  LANDSAT  TM  image. 
From  these  GCPs,  the  mean  and  standard  deviation  of  the  absolute  differences  along 
with  the  2D-RMS  and  Circular  Error  probability  (CEP)  radius  are  computed. 

The  specified  positional  accuracy  of  the  woody  vegetation  dataset  at  1 : 100,000 
scale  is  90%  <  50  metres.  The  results  from  the  3  spot  checks  indicate  that  all  the 
mean  2D-RMS  values  are  >  50m  and  the  85%  CEP  are  2  to  3  times  the  50m 
specification. 


Budgetary  constraints  during  the  implementation  of  the  accuracy  assessment  methodology  dictated  that  the 
sampling  grids  be  reduced  to  10  x  15  primary  and  6x6  secondary  points  giving  a  total  of  5,400  sampling 
points. 


322 


SAMPLE  SIZE  DETERMINATION 


The  sampling  unit  for  the  attribute  accuracy  assessment  is  a  patch  of  woody 
vegetation.  The  minimum  Basin  wide  patch  size  is  1  ha.  The  total  number  of  1  ha. 
patches  is  N  =  105,800,000  and  at  the  minimum  mapping  unit  size  of  0.25  ha.,  N  = 
423,200,000. 

One  of  the  statistical  problems  faced  in  the  sample  design  is  that  there  are  very 
few  accepted  guidelines  for  determining  sample  sizes  in  spatial  analysis.  Many  well 
respected  authors  and  studies  pluck  a  sample  size  from  the  ether  with  naught 
justification.  Goodchild  et.al.  (1991)  uses  a  sample  size  of  1347  sites  systematically 
sampled  from  56,973  sites  and  does  not  described  how  they  arrived  at  this  figure. 

The  beginnings  of  the  statistical  definition  of  sample  size  in  spatial  analysis  is 
seen  in  the  early  work  of  Van  Genderen  et.al.  (1978).  They  outline  the  use  of  the 
binomial  distribution  as  a  means  for  deciding  the  sample  size  based  on  a  confidence 
requirement  (95%)  and  a  specified  minimum  classification  error  in  the  population 
of  85%  correct.  This  corresponds  with  the  US  Geological  Survey  Circular  671 
operational  job  specification.  Rosenfield  &  Melly  (1980)  and  Veregin  (1989) 
provide  a  similar  rationale  based  on  the  confidence  interval  for  proportions,  again 
based  on  the  binomial  distribution. 

At  the  secondary  grid  level,  the  proposed  sample  design  takes  an  area  of 
10x1 0km.  This  corresponds  to  10,000  x  1  ha.  patches.  Thus  based  on  the  binomial 
theory,  a  sample  of  60  one  hectare  patches  from  10,000  one  hectare  patches,  a 
sample  fraction  of  0.6%,  should  provide  an  unbiased  estimate  of  the  population  (here 
a  10  x  10km  area)  classification  accuracy. 

However,  at  the  Basin  wide  level,  60  one  hectare  sample  points  amongst 
105,800,000  (a  sample  fraction  of  0.0001%)  is  a  tad  small!  Theoretically  it  is 
justifiable.  Typical  telephone  polling  of  the  entire  Australian  population  has  sample 
fractions  of  0.01%,  100  times  that  of  the  above! 

Unfortunately,  the  Remote  Sensing  &  GIS  literature  gives  little  guidance  as  to 
what  constitutes  an  acceptable  sample  size.  Discussions  with  Dr.  Ray  Czaplewski 
and  Gretchen  Moisen  confirmed  this  view.  Congalton  (1991)  offers  a  rule  of  thumb 
of  100  sample  sites  per  classification  category.  Czaplewski  &  Catts  (1992) 
recommend  sample  sizes  of  500  to  1000  based  on  simulation  studies. 

The  decision  on  the  final  sample  sizes  was  constrained  by  3  factors.  The  number 
of  primary  sampling  points  needed  to  be  as  large  as  practicable  to  minimise  the 
contour  interpolation  error.  The  secondary  grid  sample  size  of  60  (95%  correct 
classification  with  95%  confidence)  based  on  the  binomial  distribution  seemed 
reasonable  based  on  the  literature  and  discussions  with  colleagues.  The  sample  size 
had  to  accommodate  the  patch  size  sampling  frames  on  the  transparent  overlay. 

The  compromise  decided  on  was  a  primary  grid  of  12x18  (216  points)  and  a 


323 


secondary  grid  of '7x7  (49  points)  giving  a  total  sample  size  of  10,584  patches.  The 
secondary  grid  size  is  statistically  defensible  and  the  primary  grid  size  gives 
sufficient  control  points  for  the  contour  interpolation. 

ANALYSIS  OF  THE  ATTRIBUTE  AND  POSITIONAL  ACCURACY. 

The  Positional  accuracy  assessment  will  be  based  on  a  summary  table  of  the 
absolute  differences  between  the  map  and  digital  image  coordinates  (E  &  N)  of  the 
GCPs  recorded  at  the  secondary  grid  level  for  each  primary  grid  point.  The  GCP 
outliers  (extreme  differences  i.e.  >  1km),  will  be  identified,  documented  and 
removed  from  the  analysis.  Summary  statistics  (means,  standard  deviations, 
maximum  &  minimum)  for  each  primary  grid  point  will  be  aggregated  into  a  Basin 
wide  positional  accuracy  report.  The  2D-RMS  and  Circular  error  probability  radius 
(CEP)  for  each  primary  grid  point  will  be  created  from  the  raw  GCP  differences 
discussed  above.  A  contour  plot  of  the  2D-RMS  from  each  primary  grid  point  will 
be  produced.  This  contour  map  illustrates  the  spatial  variability  of  positional 
accuracy  across  the  MDB. 

The  Attribute  accuracy  assessment  follows  the  trend  in  most  of  the  literature  in 
utilising  an  analysis  of  the  error  matrix.  For  each  Primary  grid  point  and  for  each 
patch  size,  an  error  matrix  of  the  pseudo  ground  truth  values  of  woody/  non  woody 
derived  from  the  air  photo  interpretation  versus  the  values  derived  from  the  digital 
imagery  is  constructed. 

From  this  error  matrix  (for  each  patch  size)  the  Overall  Accuracy  %  (OA%), 
Kappa  statistic  and  user  and  producer  accuracies  are  computed  for  each  primary  grid 
point  and  for  each  patch  size.  They  can  be  reported  separately  as  the  secondary  level 
sample  size  is  sufficient  to  make  them  statistically  valid.  The  contents  of  the  Basin 
wide  attribute  accuracy  summary  table  is  then  contour  mapped  to  produce  the  spatial 
variability  maps  of  Attribute  accuracy. 

CONCLUSIONS 

The  attribute  accuracy  assessment  project  developed  and  trialed  a  methodology 
for  assessing  the  attribute  and  positional  mapping  accuracy  of  the  woody  vegetation 
layer  (woody/  non-woody)  within  the  MDB  dataset.  The  flexibility  of  two  stage 
systematic  sampling,  its  simplicity  of  implementation,  the  large  number  of  regular 
grid  points  distributed  across  the  basin  at  different  sampling  levels  and  the  potential 
flexibility  for  post  stratification  and  interpolation  were  the  deciding  factor  in 
choosing  it  for  the  attribute  accuracy  assessment  for  the  MDBP. 

ACKNOWLEDGEMENTS 

I'm  indebted  to  the  practical  experience  of  Adam  Choma  of  CNR,  Peter  Knock 


324 


of  DLWC,  Bathurst  and  Graeme  Dudgeon  (NSW  Dept.  Agriculture,  Orange).  Mrs. 
Kim  Smith  my  research  assistant  persevered  with  the  GCPs.  On  the  statistical  front. 
Dr.  Ray  Czaplewski  of  the  US  Forestry  Service,  Fort  Collins  Colorado  USA  proved 
an  invaluable  sounding  board.  Gretchen  Moisen  of  the  US  Dept.  Of  Agriculture, 
Ogden  Utah  USA  provided  insights  into  accuracy  assessment  on  large  scale  datasets. 
Dr.  Brian  Lees  (Australian  National  University,  Geography  Dept.)  provided  early 
thoughts. 

REFERENCES 

Cochran.  W.G.,  1977,  Sampling  Techniques  (3rd  ed.),  John  Wiley  &  Sons,  NY 
Congalton,  R.G.,   1988,  Comparison  of  sampling  schemes  used  in  generating 
error  matrices  for  assessing  the  accuracy  of  maps  generated  from  remotely 
sensed  data. ,  Photo grammetric  Engineering  and  Remote  Sensing,  v  54,  n  5  May 
1988,  p  593-600 

Congalton  R.G.,  1991,  A  review  of  assessing  the  accuracy  of  classifications  of 

remotely  sensed  data.,  Remote  sensing  of  Environment,  v  37,  n  1,  p  35-46 
Crapper,  P.F.,  Walker,  P. A.,  Nanninga,  P.M.,  1986,  Theoretical  prediction  of  the 

effect  of  aggregation  on  grid  cell  data  sets.,  Ge-processing,  3,  pi  55-166. 
Czaplewski,  R.  &  Carts,  G.P.,  1992,  Calibration  of  Remotely  Sensed  proportion  or 

area  estimates  for  misclassification  error.  Remote  Sensing  of  Environment,  v  39, 

p  29-43 

Czaplewski,  R.,  Goodson,  P.,  Gilbee,  A.,  Razier,  P.,  Choma,  A.,  1992,  Accuracy 
assessment  of  Remotely  Sensed  thematic  maps.  Project  Brief,  Dept  Conservation 
and  Environment,  Victoria,  Australia,  September  29,  1992. 

Fitzgerald,  R.W.  &  Lees,  B.G.,  1994a,  Assessing  the  classification  accuracy  of 
multisource  remote  sensing  data.,  Remote  Sensing  of  Environment,  47, n  3, 
362-368 

Fitzgerald,  R.W.,  &  Lees,  B.G.,  1994b,  Spatial  Context  and  scale  relationships 
in  raster  data  for  thematic  mapping  in  natural  systems,  Spatial  Data  Handling 
Conference,  Edinburgh,  Scotland,  Sept  1994.  Published  by  Taylor  &  Francis,  UK 
(in  press). 

Franklin,  J.,  Hiernaux,  P.H.Y.,  1991,  Estimating  Foliage  and  Woody  Biomass 
in  Sahelian  and  Sudanian  Woodlands  Using  a  Remote-Sensing  Model., 
International  Journal  of  Remote  Sensing,  v  12,  n  6,  p  1387-1404 

Goodchild,  M.F.,  Davis,  F.W.,  Painho,  M.,  Stoms,  D.M.,  1991.  The  use  of  vegetation 
maps  in  geographic  information  systems  for  assessing  conifer  lands  in  California. 
National  Centre  for  Geographic  Information  and  analysis,  Dept  Geography,  Uni 
California.  Report  91-2B  NCGIA 

Gilbee.  A.  &  Goodson,  P..  1992,  Mapping  tree  cover  across  Victoria  using 
Thematic  Mapper  digital  data  and  a  GAS.,  Proc.  6th.  Australian  Remote 


325 


Sensing  Conference,  Wellington,  NZ,  Nov  1992. 
Janssen,  L.L.F.  &  van  der  Wei,  F.J.M.,  1994,  Accuracy  assessment  of  satellite 

derived  land-cover  data:  a  review,  Photo grammetric  Engineering  &  Remote 

Sensing,  v  60, n  4,  April,  p  419-426 
McDonald,  R.C.,  Isbell,  R.F.,  Speight,  J.G.,  Walker,  J.  &  Hopkins,  M.S.,  1990, 

Australian  Soil  and  Land  Survey;  Field  Handbook,  ed.  2,  Inkata  Press, 

Melbourne,  Australia. 
Moisen,  G.G.,  Edwards,  T.C.  Jnr,  Cutler,  D.R.,  1994,  Spatial  sampling  to  assess 

classification  accuracy  of  Remotely  Sensed  data.  Environmental  Information 

Management  and  Analysis:  Ecosystem  to  Global  Scales  Michener,  Brunt  and 

Stafford  (eds).  p  159-176 
Murray  Darling  Basin  Commission,  1995,  Recipe  for  the  Attribute  and  Positional 

Accuracy  Assessment  of  the  Murray  Darling  Basin  Project.,  internal  working 

document,  by  R  W  Fitzgerald,  Infoplus  for  DLWC  Land  Information  Centre, 

Bathurst,  Australia. 

Ritman,  K.T.,  1995,  Structural  vegetation  data;  A  specifications  for  the  Murray 
Darling  Basin  Project  M305.  DLWC,  Land  Information  Centre,  Bathurst, 
Australia 

Rosenfield,  G.H.  &  Fitzpatrick-Lins,  K.,  1986,  A  coefficient  of  agreement  as  a 

measure  of  thematic  classification  accuracy.,  Photogrammetric  Engineering  and 

Remote  Sensing,  v  52,  n  2,  p  223-227. 
Rosenfield,  G.H.  &  Melly,  M.L.,  1980,  Applications  of  statistics  to  thematic 

mapping,  Photogrammetric  Engineering  and  Remote  Sensing,  v  46,  n  10, 

October,  p  1287-1294 
Tadrowski,  T.,  Hart,  D.G.D.,  Schepp,  K.,  1990,  A  study  of  the  possibilities  and 

accuracies  of  1:50  000  vegetation  mapping  using  Remotely  Sensed  data.  8th 

Australian  Inst  Cartographers  Conference,  Darwin,  Australia,  May  1990. 
Van  Genderen,  J.L.,  Lock,  B.F.  &  Vass,  P. A.,  1978,  Remote  Sensing:  Statistical 

testing  of  thematic  map  accuracy.  Remote  Sensing  of  the  Environment,  v  7,  p  3- 

14. 

Veregin,  H.,  1989,  A  taxonomy  of  error  in  spatial  databases.  National  Centre  of 
Geographic  Information  and  Analysis,  Technical  Report  89-12,  December  1989. 


326 


A  New  Method  for  Evaluating 
Positional  Map  Accuracy 

Michal  Lodin1  and  David  Skea2. 


Abstract. -Typically,  positional  map  accuracy  is  determined  from  a  small 
number  of 'well  defined  points'.  If  a  sufficient  number  of  such  points  exist, 
it  is  assumed  that  they  are  evenly  distributed  over  the  map,  and  represent 
the  overall  map  accuracy.  In  this  study,  a  method  of  direct  examination  of 
the  geometry  representing  point,  linear,  and  aerial  features  is  described. 
Pattern  recognition  procedures  are  used  to  automatically  extract  and 
match  all  identifiable  nodes,  arcs,  and  polygons  of  an  accurate  benchmark 
map,  and  a  test  map.  Different  measures  of  "similarity"  are  computed  for 
corresponding  map  elements,  -  cut  blocks  and  logging  roads.  Due  to  the 
automated  approach,  large  datasets  containing  virtually  all  matching 
features  are  compiled.  Thus,  distribution,  correlation,  and  spatial 
variation:  tests,  typically  not  done  in  standard  accuracy  assessments,  can 
be  earned  out.  Various  mapping  methods,  operationally  used  in 
maintaining  the  inventory  map  database  of  the  British  Columbia  Forest 
Service  were  tested  using  the  developed  methodology. 


INTRODUCTION 

The  British  Columbia  Forest  Service  manages  well  over  six  thousand  1:20  000 
mapsheets  covering  an  area  of  approximately  83  million  hectares.  Forest  openings 
resulting  from  timber  harvesting  and  road  construction  are  updated  on  a  two  year 
cycle.  The  map  update  process  is  carried  out  both  by  Forest  Service  staff  and 
through  contractors.  The  objective  of  this  study  was  to  develop  and  test  a  suitable 
methodology  to  evaluate  positional  accuracy  of  map  update  procedures  used  in  the 
Forest  Inventory  program.  Maps  obtained  by  operationally  available  methods  were 
compared  to  a  highly  accurate  benchmark  through  a  set  of  objective  procedures 
which  provide  a  quantitative  measure  of  their  mutual  "dissimilarity". 

Positional  mapping  accuracy  standards,  as  adopted  by  most  modern  mapping 
programs,  state  that  a  percentage  of  all  well-defined  planimetric  features  are 
located  within  some  minimum  distance  of  their  geographic  position  with  reference 
to  a  prescribed  datum  (Committee  for  Standards  and  Specifications,  1985).  To  test 


^Remote  Sensing  Specialist.  Ministry-  of  Forests.  Victoria.  British  Columbia,  Canada 

"  Systems  Analyst.  Ministry  of  Environment.  Lands  and  Parks.  Victoria.  British  Columbia.  Canada 


327 


if  a  given  map  meets  a  stated  standard,  the  location  of  a  number  of  well-defined 
points  are  sampled  and  compared  with  their  true  locations,  as  determined  by  some 
highly  accurate  positioning  method  (e.g.,  differential  GPS).  The  differences 
between  the  two  positions  are  then  computed  and  summary  statistics  such  as  RMS 
and  CSE  (Circular  Map  Error)  are  computed.  An  outline  of  this  methodology  can 
be  found  in  Maling  (1989)  or  Merchant  (1987).  Another  example  ,  applied  to 
forest  mapping,  is  given  in  DMG  (1991).  Unbiased  selection  of  well  defined  points 
is  another  problem.  Dozier  and  Strahler  (1983)  have  suggested  that  all  possible 
well  defined  points  be  extracted  from  the  map  and  then  from  these  a  random 
selection  of  these  points  be  used  as  the  sample.  Lack  of  resources  to  collect 
sufficient  ground  information  is  common  (Fenstermaker,  1991). 


METHODOLOGY 


Both  test  sites  were  mapped  using  the  most  accurate  method  available.  The 
resulting  benchmark  map  is  of  a  higher  order  of  accuracy  than  that  produced  by 
the  tested  methodologies.  The  same  site  is  then  remapped  using  the  test 
methodologies,  and  these  new  maps  are  compared  to  the  reference  map.  From  the 
observed  differences,  a  statistical  picture  of  the  similarity  of  the  new  map 
compared  to  the  reference  map  is  derived.  All  tests  are  empirical,  thus  no  effort  has 
been  made  to  separate  different  error  components.  In  teims  of  GIS  error  taxonomy 
(Collins  and  Smith,  1994),  most  errors  described  here  can  be  assumed  to  be 
category  E12)5,6,p,s,e- (systematic  and  random  positional  errors  due  to  data 
collection,  input,  output  and  interpretation).  A  diagram  showing  the  developed 
procedure  is  given  in  figure  1.  If  the  reference  map  is  considered  error-free,  this 
statistical  picture  would  reflect  the  accuracy  of  the  new  maps. 


test  map 


nodes 
arcs 

polygons 


benchmark 


nodes 
arcs 

polygons 


node  offsets 
arc  offsets 

polygon  boundary  offsets 
polygon  area  differences 
polygon  slivers 
arc  shape  differences 
polygon  shape  differences 

node-node 
arc-art 

polygon-polygon 

 ► 

feature  extraction 


feature  matching 


difference  observations 


Figure  1.—  Process  diagram  for  automated  positional  map  accuracy  assessment. 


Benchmark  map  compilation 


Two  test  sites  were  selected,  both  with  significant  relief  and  both  containing  a 
large  number  of  features  to  be  mapped.  High  accuracy  (as  compared  to  the  tested 


328 


methods)  benchmark  maps  were  compiled  for  both  sites  using  GPS  ground  control 
points  and  analytical  plotter  instrument.  Positional  accuracy  of  one  of  the 
benchmarks  was  evaluated  using  seventeen  'well  defined  points'  to  be  within  RMS 
7.35  meters  (X,Y)  from  their  GPS  field  surveyed  positions. 

Feature  extraction 

Benchmark  and  test  map  were  compiled  in  an  unstructured  format  (noded  and 
cleaned  result  of  'spaghetti  digitizing').  Basic  geometric  objects:  a)  polygons 
representing  forest  openings,  b)  arcs  constituting  roads  and  partial  polygon 
boundaries,  and  c)  nodes  identifying  road-road  and  road-forest  opening 
intersections  were  formed.  Subsequently,  an  internal  data  structure,  the  kd  tree 
was  constructed,  (Ooi,  1990).  Each  node  of  the  tree  contains  one  point  and  divides 
its  subdivision  of  the  plane  into  two  parts  (Fig.  2,  3).  End  points  of  each  line 
segment  contain  links  defining  the  line  segment  and  the  left-to-right  sorting  of  the 
nodes.  Each  time  an  end  point  is  found  to  lie  within  a  given  tolerance  of  an  existing 
tree  node,  the  order  of  the  tree  node  is  increased.  Hence,  nodes  of  order  one 
represent  line-ends  not  connected  to  another  feature.  Nodes  of  order  two  are 
simple  two-line  connections.  Nodes  of  order  three  or  higher  represent  line 
intersections.  The  built  segment  tree  is  scanned  to  extract  nodes  arcs  and  polygons. 


(0,100) 


q  D(30,90) 


B(10.75) 
 ©- 


(0.0) ' 


<&  C(25,15) 


()F(80,70) 


()  A(40,60) 


(100,100) 


E(70,20) 


(100,0) 


Figure  2.— Points  on  the  plane. 


Figure  3.— The  resulting  kd-tree. 


Feature  matching 

Automated  feature  extraction,  matching,  and  differencing  procedures  have  been 
developed  on  the  MicroStation  platform  as  MDL  commands.  The  matching 
process  was  partially  hampered  by  the  interpretive  nature  of  the  features  involved. 
Polygon  boundaries  interpreted  from  LANDSAT  imagery  did  not,  in  some  cases, 


329 


resemble  the  'true'  extent  of  that  polygon  on  the  benchmark  map  (Fig.  4)  and  some 
logging  roads  were  simply  not  discernible.  Matching  of  extracted  geometric 
objects  is  facilitated  by  computing  specific  attributes  for  each  object  and  defining 
some  type  of 'distance'  measure  (see  Haralick  and  Shapiro,  1992).  An  example  is 
given  below: 

Nodes  were  matched  based  on  position,  number  of  lines  making  up  the  intersection,  and 
relative  angle  between  the  lines.  For  each  node  on  the  benchmark  map,  a  range  rectangle  is 
defined.  Each  point  on  the  benchmark  which  lies  within  this  range  rectangle  is  tested.  If  the 
benchmark  node  has  the  wrong  number  of  lines  then  it  is  rejected.  For  the  remaining  nodes, 
a  weighted  measure  based  on  the  position  difference  and  relative  angles  is  computed  and 
the  'closest'  node  is  accepted  as  the  match. 

Arcs  and  polygons  were  matched  using  similar  techniques.  Polygons  matching 
was  close  to  100%  effective.  Node  and  arc  matching  is  in  the  order  of  80% 
effective.  To  correct  miss-matches,  procedures  to  interactively  view  and  correct 
node,  arc,  and  polygon  matches  have  been  developed. 

Computed  difference  measures  were: 
Node  offsets  -  (  ax  ,  a  y  )  in  meters, 

Arc  and  polygon  closest  point  offsets  -  Offset  of  each  arc/polygon  vertex  from  the  closest 
vertex  of  the  matching  arc/polygon. 

Arc  and  polygon  perpendicular  bisectors  -  Perpendicular  bisector  from  each  line 
connecting  two  vertices  to  the  matching  line  segment  of  the  corresponding  arc/polygon  line 
segment. 

Polygon  area  differences  -  Area  and  perimeter  differences  between  matched  polygons. 
Polygon  slivers  -  Two  corresponding  boundaries  of  matching  polygons  typically  cross 
each  other  a  number  of  times.  These  crossings  create  sliver  polygons.  For  each  matched 
polygon,  slivers  are  extracted  and  their  centroid  position,  area  and  perimeter. 


Figure  4.— Example  of  matched  polygons  (70)  and  polygons  rejected  due  to  interpretation 
differences  (73). 

Tested  methods  and  compiled  datasets 


330 


Table  1  shows  the  fourteen  mapping  systems  tested.  The  result  of  all  mapping  was 
a  digital  file  in  IGDS  format  For  analog  systems,  such  as  PROCOM  II,  this 
required  an  additional  manual  digitization  step  (consistent  with  normal  practice). 


Table  1. —Tested  methodologies. 


Source  data 

Format 

Mapping  method 

Contract  or 

in-house 

production 

LANDSAT  TM*  . 

1:1  000  000,  1:500 
000 

Analogue, 
transparency* 

PROCOM  II 

Contract 

SPOT  PAN**,  1:500 

AAA       1     ^CA  AAA 

000.  1 :2>0  000 

Analogue, 
transparency* 

PROCOM  II 

Contract 

LANDSAT  TM 
^  i  neniauc  iviapper^ 

Digital,  various  levels 
or  geomeinc 
corrections 

Digital  update 

In-house 

SPOT  Panchromatic 

Digital 

Digital  update 

In-house 

SPOT  Multispectral 

Digital 

Digital  update 

In-house 

Aerial  photography, 
1:40  000 

Analogue 

Mono  restitution 

Contract 

Aerial  photography, 
1:40  000 

Scanned  @  853  dpi. 
1 .2m  pixel  size 

DIAP  (Digital  Image 
Analytical  Plotter) 

Contract 

Aerial  photography, 
1:40  000 

Scanned,  resampled 
to  4.8m  pixels 

DVP  (Digital  Video 
Plotter) 

In-house 

The  number  of  "correctly"  mapped  and  matched  features  varies  from  one  method 
to  another.  Typically,  lower  order  methods,  such  as  optical  projection  PROCOM 
II  represent  the  lower  bound  while  higher  order  softcopy  photogrammetric 
methods  (e.g.,  DIAP)  produced  results  of  the  upper  limit  of  the  range.  For  some 
methodologies,  the  number  of  matched  nodes,  ('well  defined  points'  9)  was  as  low 
as  10.  Large  data  sets  were  are  collected  for  arc  and  polygon  boundary  offsets  as 
well  as  for  polygon  slivers.  For  more  detail  and  test  results,  see  Skea  and  Lodin 
(1995). 


Table  2. —Size  of  compiled  datasets. 


Features  type  matched  and  evaluated 

dataset  compiled  for  one  test  area 

Node  offsets  road/road  and  road/polygon 

10-300  matched 

arc  and  polygon  boundary  shortest  distance 
and  perpendicular  offsets 

2,000  -  5,000  observations  in  each  dataset 

polygon  area,  position,  shape 

50-90  matched 

sliver  polygons  ("in"  and  "out"  relative  to 
benchmark) 

400  -  1800  -  in  each  dataset 

331 


Various  distributions  of  polygon  boundary  error  observations  have  been  reported.  Errors 
associated  with  manual  digitizing  (Bolstad  et  al  ,1990)  were  found  to  conform  to  a  Chi- 
squared  distribution. 

Automated  observation  techniques  used  in  this  study  produce  data  sets  of 
sufficient  size  to  directly  test  for  distribution  (Fig.  6).  These  observations  suggest,  that 
distribution  of  offsets  cannot  be  assumed  to  be  normal.  Thus,  estimates  of  epsilon 
bandwidth  (Chrisman,  1989),  used  for  polygon  and  arc  boundary  error  estimates,  should 
be  based  on  distribution  free  statistics. 


Figure  5.~Quantile-quantile  distribution  plot  of  polygon  offsets  collected  by  two  different 

methodologies,  softcopy  photogrammetric  device  using  1:40  000  photography  (L),  and  an 
optical  projection  PROCOM  II  system  using  LANDSAT  TM  1:500  000  transparency  R). 
Note,  in  a  qq  plot,  cluster  along  a  straight  line  indicates  normal  distribution. 

On  the  epsilon  bandwidth  plot  (Fig.  6),  the  median  value  and  epsilon  bandwidth  or  EBW 
of  displacements  of  polygon/arc  boundary  are  given.  The  median  value  shows  any  skew  in 
the  data  (ideal  data  should  be  centered  around  zero)  and  the  EBW  is  computed  as: 
+  1.4826  x  median  x  (y  -  center).  This  value  corresponds  to  one  sigma  (the  standard 
deviation)  for  Gaussian  data  but  is  robust  with  respect  to  distribution.  The  perimeter  of 
each  polygon  is  normalised  between  zero  and  2n. 


Polygon  Displacement 


mun=  0  2  781 
2  sigmi=  28  'l  72 


Kamioom  MONO 


Figure  6.-Epsilon  band  plot  of  polygon  displacements  for  the  Mono  restitution  method. 


332 


RESULTS 


The  focus  of  the  statistical  analysis  was  on  assessing  the  form  of  the  data  and  the 
relative  accuracy,  rather  than  just  producing  descriptive  statistics.  This  included 
testing  for  distribution,  computing  variance  of  polygon  area  differences,  and 
estimating  bandwidth  for  polygon  boundaries  and  arcs.  Standard  descriptive 
measures  such  as  circular  map  errors  were  also  computed.  Results  were  presented 
in  a  comprehensive  report  (Skea  and  Lodin,  1995)  at  different  levels  of  detail  in 
the  following  categories: 

Detection  limitations.  The  number  of  map  features  potentially  omitted  due  to  data 
resolution,  interpretation  limitations  or  simply  due  to  the  "fuzzy"  nature  of  forest 
openings. 

Positional  error  of  intersections  (nodes).  A  statistical  measure  used  for  the 
Provincial  base  mapping  program  (Circular  Map  Standard  Error)  was  found  well 
suited  for  node  accuracy  assessment.  Surprisingly,  positional  errors  of  intersections 
(nodes)  were  systematically,  for  all  tested  methods,  found  not  indicative  of  the 
overall  map  accuracy  (Fig.  8). 


Figure  7.--Example  of  a  node  offset  which  is  not  representative  of  overall  map  accuracy. 
This 'T'  shaped  node  formed  by  a  road/road  intersection  captured  from  a  LANDSAT 
satellite  image  (dashed  line)  is  compared  to  the  benchmark  node  w  hich  appears  skewed. 
Clearly,  the  resulting  offset  is  not  indicative  of  the  overall  accuracy  of  the  map. 


Positional  error  of  roads  and  cutblock  boundaries.  Epsilon  bandwidth  was  found 
useful  as  a  means  of  expressing  line  (arc  and  polygon)  accuracy.  It  was  used  to 
compare  different  methodologies  ,  in  the  context  of  this  study,  as  it  is  less 
dependent  on  distribution  of  data  compared  to  other  types  of  statistical  measures. 

Area  error  of  cutblocks.  Overlays  of  benchmark  and  test  map  polygons  produce 
sliver  polygons.  Areas  of  sliver  polygons  in-  and  out-  relative  to  the  benchmark 
were  used  as  measures  of  systematic  cutblock  area  under-  or  over-  estimate 
associated  with  a  given  mapping  method.  Comparing  polygon  areas  only  was 


333 


found  biased  as  it  would  favor  methods  capable  of  mapping  only  large  cutblock 
polygons  (e.g.,  LANDSAT). 

CONCLUSIONS 

Automated  pattern  matching  techniques  can  be  applied  successfully  to  the  problem 
of  positional  map  accuracy  assessment.  Their  use  leads  to  large  and  more  diverse 
sets  of  observations  on  which  to  judge  map  accuracy.  Most  importantly,  positional 
accuracy  of  different  types  of  map  features  cannot  be  assumed  the  same  for  a  given 
method.  This  conclusion  may  be  affected  by  nature  of  mapped  features  in  this 
study  (forest  openings).  Although  the  use  of  these  observations  clearly  requires  the 
development  of  new  statistical  analysis  procedures,  they  also  allow  for  the 
development  of  broader  statistical  descriptions  of  the  data. 

REFERENCES 

Bolstad,  P.  V.,  Gessler,  P.  and  Lillesand,  T.M.,  1990.  Positional  Uncertainty  in 
Manually  Digitized  Map  Data.  International  Journal  of  Geographical 
Information  Systems,  4  (4),  399-412. 

Chrisman,  N.R.,  1989.  Modeling  Error  in  Overlays  of  Categorical  Maps.  In, 

Accuracy  of  Spatial  Databases,  edited  by,  Goodchild,  M.  and  Gopal,  S.,  Taylor 
and  Francis. 

Collins,  F.C.  and  Smith,  L.J.,  1994.  Taxonomy  for  Error  in  GIS.  In,  Proceedings, 
International  Symposium  on  the  Spatial  Accuracy  of  Natural  Resource  Data 
Bases,  1-7.  ASPRS. 

Committee  for  Specifications  and  Standards.  1985,  Accuracy  Specification  for 
Large-Scale  Line  Maps.  Photogrammetric  Engineering  and  Remote  Sensing, 
Vol.  51,  No.  2,  195-199. 

DMG  (Digital  Mapping  Group  Ltd.),  1991.  Project  to  Evaluate  Photogrammetric 
Accuracy  of  Several  Plotting  Instruments  to  Transfer  Forest  Inventory 
Information  from  Aerial  Photographs  to  TRIM  Base  Maps.  Technical  Report, 
British  Columbia  Ministry  of  Forests,  Inventory  Branch. 

Dozier,  J.  and  Strahler,  A.H..  1983.  Ground  Investigations  in  Support  of  Remote 
Sensing.  Chapter  23,  Manual  of  Remote  Sensing,  second  edition,  Simonett, 
D.S.  and  Ulaby,  F.T.  editors,  American  Society  of  Photogrammetry  and 
Remote  Sensing. 


334 


Fenstermaker,  L.K  ,  1991.  A  Proposed  Approach  for  National  to  Global  Scale 
Error  Assessments.  In,  Proceedings,  GIS/LIS  '91,  Vol.  1,  293-300. 

Haralick,  R.M.  and  Shapiro,  L.G.,  1992.  Computer  and  Robot  Vision,  Volumes  I 
&  II.  Addison-Wesley  Publishing  Company. 

Maling,  D.H.,  1989.  Measurements  from  Maps:  Principles  and  Methods  of 
Cartometry.  Pergamon  Press. 

Merchant,  DC,  1987.  Spatial  Accuracy  Specification  for  Large  Scale 

Topographic  Maps.  Photogrammetric  Engineering  and  Remote  Sensing  53  (7): 
958-61. 

Ooi,  B.C.,  1990.  Efficient  Query  Processing  in  Geographic  Information  Systems. 
Goos,  G.  and  Hartmanis,  J.  (ed  ),  Lecture  Notes  in  Computer  Science,  No.  471, 
Springer- Verlag. 

Skea,  D.  and  Lodin,  M.,  1995,  Computer- Assisted  Methods  to  Assess  the 
Positional  Accuracy  of  Map  Overlays.  Ministry  of  Forests,  Resources  Inventory 
Branch,  Technical  report. 

BIOGRAPHICAL  SKETCH 

Michal  Lodin  holds  Masters  degrees  in  Geology  from  Charles  University, 
Prague  and  Surveying  Engineering  from  the  University  of  New  Brunswick.  After  8 
years  of  consulting  in  Geomatics  applications  in  Canada,  U.S.  and  Europe,  he  now 
works  in  the  Inventory  Technical  Support  Section,  Resources  Inventory  Branch, 
Ministry  of  Forests. 

David  Skea's  photogrammetric  training  includes  diplomas  from  the  British 
Columbia  Institute  of  Technology  and  the  International  Institute  for  Aerial  Survey 
and  Earth  Sciences  in  Holland.  His  research  interest  are  in  areas  of  pattern 
recognition,  computer  vision  and  computational  geometry.  He  is  employed  with 
Geographic  Data  B.C.  heading  a  project  to  develop  an  automated  watershed  and 
stream  network  mapping  system. 


335 


Exploring  Spatial  Confidence  with  a 

Raster-based  GIS 


Matthew  H.  Pelkki1 


Abstract.--  One  of  the  main  input  methods  into  geographic  information 
systems  remains  digitizing  spatial  data  from  paper  maps.  While  newer, 
automated  methods  such  as  global  positioning  systems  promise  better 
accuracy  and  quality  of  data  for  future  GIS,  digitizers  are  likely  to  be  part  of 
the  data  collection  technology  for  some  time  into  the  future.  When  digitizing 
this  data,  it  is  important  to  remember  that  a  single  sample  of  the  representation 
of  reality  is  being  taken.  Information  related  to  the  spatial  confidence  of  the 
resulting  representation  often  is  not  quantified  well  by  the  digitizing  software. 
Spatial  confidence  can  be  expressed  as  percentage  or  degree  of  certainty  that 
a  represented  object  does  indeed  exist  at  that  location  or  within  a  certain 
distance  of  that  location.  This  research  found  that  for  map  registration  error 
may  not  be  a  good  indication  the  required  raster  resolution  to  obtain  high 
spatial  confidence.  It  was  also  found  that  complex  lines  tend  to  increase  the 
amount  of  error  or  variance  in  the  resulting  digital  map. 


INTRODUCTION 

Obtaining  spatial  data  of  high  quality  is  an  obvious  desire  of  spatial  data  users. 
Positional  accuracy  is  one  important  component  of  data  quality  (Antenucci  et  al, 
1991).  Error  in  spatial  data  is  inherent  and  difficult  to  remove.  While  minimizing 
error  has  long  been  a  goal  of  spatial  sciences  (Chrisman,  1991),  others  (Aronoff 
1989)  suggest  that  since  error  cannot  be  eliminated,  it  should  be  managed.  Certainly, 
the  first  step  in  managing  error  is  being  able  to  quantify  it. 

The  need  to  explicitly  understand  spatial  confidence  in  a  GIS  is  growing  as  digital 
data  becomes  less  specialized  and  more  widespread  and  mainstream.  It  is  well  known 
that  digital  data  is  perceived  to  be  of  higher  quality  than  analog  data  due  to  the 
appearance  of  precision  and  accuracy.  As  developers  of  digital  data  become  more  and 
more  removed  from  the  end-users,  explicit  understanding  of  locational  confidence  is 
important  to  prevent  mis-application  of  data  to  analyses  for  which  it  is  not  suitable. 
Paper  maps  are  currently  the  most  common  source  for  geographic  data,  and  while 


Assistant  Professor,  University  of  Kentucky,  Department  of  Forestry,  Lexington,  KY  40546-0073. 


336 


automated  methods  such  as  global  positioning  systems  promise  better  accuracy  and 
quality  of  data  for  future  GIS,  digitizers  are  likely  to  be  a  part  of  the  data  collection 
technology  for  some  time  into  the  future.  Historical  data,  and  much  natural  resource 
data  will  be  available  only  on  paper  maps.  In  these  cases,  the  map  is  in  fact  the  only 
representation  of  "truth"  and  any  one  manual  digitization  of  that  map  is  but  one 
sample  of  reality.  The  map  scale  is  often  used  as  an  estimate  for  the  accuracy  of  the 
digitized  data  (Fisher,  1991)  and  one  sample  of  reality  is  often  all  that  is  taken  due  to 
time  and  cost  constraints.  This  limits  the  information  on  the  quality  of  the  data,  as  it 
assumes  that  the  first  sample  is  truly  representative  of  the  "truth." 

Ideally,  spatial  confidence  metadata  should  be  attached  to  a  raster  GIS  that  states 
something  like,  "there  is  a  95%  confidence  that  any  object  classed  within  a  raster  cell 
actually  exists  within  that  cell  or  within  a  radius  of  X  cells  from  the  indicated  cell." 
From  a  single  sample,  even  assuming  that  the  analog  data  represents  "truth,"  this  is 
difficult  to  measure.  This  research  involves  exploration  of  absolute  locational 
accuracy  of  point  and  line  data  that  is  manually  digitized  and  rasterized  to  various 
resolutions  under  the  assumption  that  the  analog  medium  being  digitized  is  in  fact 
"truth." 


METHODS 

The  GIS  software  chosen  for  this  study  was  EPPL7  (Environmental  Programming 
and  Planning  Language)  release  2.1.  This  is  a  relatively  simple  raster-based  GIS 
created  and  maintained  by  the  State  of  Minnesota  Land  Management  Information 
Center.  It  is  used  as  a  "desktop"  GIS  in  Minnesota,  and  has  users  throughout  the 
United  States  and  in  15  other  countries.  Users  include  universities,  several  state 
natural  resource  agencies,  the  National  Park  Service,  and  the  Fish  and  Wildlife 
Service. 

Two  separate  spatial  representations  were  created  digitally.  One  representation  was 
composed  of  ten  randomly  located  points.  The  second  spatial  representation  was 
composed  of  ten  lines,  two  each  having  one  to  five  line  segments.  This  was  done  to 
simulate  a  some  increasing  complexity  in  lines,  but  to  keep  it  on  a  quantifiable  scale. 
Thus,  according  to  Burrough  (1986),  more  complex  lines  (those  with  more  vertices) 
should  have  a  greater  degree  of  error  associated  with  them.  These  digital 
representations  were  considered  to  be  "truth,"  and  were  printed  out  on  a  relatively 
stable  medium  (transparency  film)  for  manual  digitizing.  The  scale  of  these  printed 
images  was  1:50,000,  and  the  point  and  line  width  printed  on  the  map  was  1/1 00th 
of  an  inch,  giving  the  approximate  representation  width  of  <  13  m. 

Both  the  point  and  the  line  images  were  digitized  ten  times  under  two  conditions. 
The  first  set  of  ten  samples  were  collected  under  a  single  map  registration,  and  the 
second  set  of  ten  samples  were  each  collected  under  a  different  map  registration.  This 
was  done  to  see  how  map  registration  differences  affected  positional  accuracy.  The 
registration  standard  deviation  of  error  was  recorded  in  each  and  every  case.  In  the 


337 


EPPL7  manual,  it  is  recommended  that  the  standard  deviation  be  less  than  the  desired 
resolution  for  rasterizing  the  file  (LMIC,  1992).  All  digitizing  was  done  in  the  center 
of  a  24  x  36  inch  Calcomp  3300  digitizing  tablet.  The  room  was  climate  controlled 
to  reduce  climate-caused  distortion  of  the  transparencies. 

Once  digitized,  the  data  files  were  rasterized  to  various  resolutions  (5,  10,  20,  and 
40  meters).  The  digital  "truth"  file  was  also  rasterized  to  the  same  resolution.  Error 
in  absolute  position  was  determined  by  overlay,  and  the  error  was  weighted  by  the 
distance  between  any  one  raster  cell  and  the  "truth." 

Similar  to  the  study  reported  by  Maffini  et  al  (1989),  this  study  examines  positional 
errors.  It  examines  the  performance  of  a  single  digitizer  operator  over  repeated  trials 
and  does  not  control  for  error  properties  in  converting  the  digital  "truth"  to 
transparency  medium,  errors  in  the  digitizing  tablet  or  equipment,  nor  control  for 
speed  of  the  operator. 

RESULTS 

Table  1  shows  the  results  from  ten  separate  digitizing  operations  on  the  ten  point 
locations  using  the  same  map  registration.  The  distances  for  locational  error  were 
calculated  by  counting  the  number  of  raster  cells  the  sample  point  was  displaced  from 
the  "truth"  point  and  multiplying  by  the  resolution  to  obtain  X  and  Y  positional  error. 
Since  under  each  resolution  there  where  ten  points  digitized  ten  times,  the  five  points 
with  the  largest  positional  errors  were  discarded,  and  the  positional  error  that  would 
contain  95  points  is  recorded  in  the  leftmost  column  of  table  1 .  We  can  see  that  as 
raster  resolution  gets  smaller,  the  95%  zone  of  inclusion  approaches  a  value  around 
20  m. 

Point  data  digitized  under  different  map  registrations  is  shown  in  table  2.  The 
rasterized  data  for  40m  resolution  were  omitted,  but  a  trend  similar  to  table  1  appears 


Table  1.  Positional  error  of  point  data  digitized  under  same  registration  and  rasterized  from 
5m  to  40m  resolution. 


Raster  cell 
resolution 

Mean  Distance 
Positional  Error 

Minimum 
Positional  Error 

Maximum 
Positional  Error 

95/100  Cells  lie 
within: 

5  m 

15.8m 

5  m 

25  m 

21.2m 

10m 

16.8  m 

0m 

28.3  m 

22.3  m 

20  m 

14.7  m 

0m 

28.3  m 

28.3  m 

30  m 

21.6m 

0m 

42.4  m 

30  m 

40  m 

18.1  m 

0m 

56.6  m 

40  m 

1  -  Std.  Dev.  of  registration  =  3. 16  m 


338 


Table  2.  Positional  error  of  point  data  digitized  under  different  registrations  and  rasterized 
from  5m  to  30m  resolution.  

Raster  cell  Mean  Distance  Minimum  Maximum  95/100  Cells  lie 


resolution  Positional  Error      Positional  Error      Positional  Error  within: 


5  m 

15.2m 

0m 

28.3  m 

25.0  m 

10m 

15.1  m 

0m 

31.6m 

28.3  m 

20  m 

13.0m 

0m 

28.3  m 

28.3  m 

30  m 

20.8  m 

0m 

42.4  m 

42.4  m 

1  -  Std.  Dev.  of  registrations  ranged  from  3.69m  to  13.17m,  and  averaged  6.93m 


to  occur.  As  raster  resolution  gets  smaller,  a  more  accurate  measurement  of  absolute 
position  error  is  possible,  and  the  maximum  value  which  includes  95%  of  all  points 
approaches  some  value. 

Determining  positional  error  for  line  data  is  a  bit  more  complex.  Error  for  lines 
includes  overshoots  and  undershoots,  as  well  as  locational  deviation  for  portions  of 
the  line.  A  line  error  index  was  calculated  for  each  digitized  line  as  it  was  compared 
to  the  "truth"  line.  The  error  index  weighted  errors  by  the  magnitude  of  the  deviation 
from  the  "truth"  line.  The  formula  used  is  as  follows: 

TV 

£  rd^Abs[N'-N]r 

LEI=—  

Nr 

where:  LEI  =  location  error  index 

N  =  number  of  raster  cells  in  "truth"  line 
N'  =  number  of  raster  cells  in  digitized  line 
r  =  raster  cell  resolution 

dj  =  distance  between  truth  line  and  digitized  line  for  cell  I 

This  index  is  independent  of  line  distance,  but  it  is  not  independent  of  resolution,  since 
with  larger  raster  cell  sizes,  small  deviations  between  the  "truth"  line  and  the  digitized 
line  will  be  impossible  to  detect.  Therefore,  the  error  index  should  decrease  with 
raster  cell  size  but  increase  with  line  complexity.  An  LEI  =  0.50  means  that,  for  any 
given  line  length,  50%  of  that  line  lies  in  a  incorrect  cell  location  weighted  by  the 
magnitude  of  the  error. 

Tables  3  and  4  show  the  LEI  value  for  lines  rasterized  to  various  resolutions  under 
the  same  and  multiple  map  registrations,  respectively.  The  tables  also  show  the 
maximum  deviation  in  number  of  cells  for  the  ten  sample  lines  from  the  "truth"  line. 
From  the  tables,  two  relationships  appear.  The  first  is  that  as  raster  cell  size  increases, 


339 


Table  3.  Line  data  digitized  under  the  same  registration  and  rasterized  from  Sm  to  40m 
resolution  (#  cells  max.  deviation  in  parentheses).  


Raster  cell  Number  of  line  segments  per  line 

resolution 

1  2  3  4  5 


5  m 

1.13(3) 

1.57  (4) 

1.51  (7) 

1.66  (5) 

2.19(6) 

10m 

0.64  (2) 

0.80  (2) 

0.78  (3) 

0.83  (3) 

1.10(3) 

20  m 

0.39(1) 

0.43  (1) 

0.47(1) 

0.48  (2) 

0.61  (2) 

30  m 

0.17(1) 

0.32(1) 

0.30(1) 

0.33(1) 

0.46(1) 

40  m 

0.11(1) 

0.23(1) 

0.26(1) 

0.23(1) 

0.35(1) 

Table  4.  Line  data  digitized  under  different  registrations  and  rasterized  from  Sm  to  40m 
resolution  (#  cells  max.  deviation  in  parentheses). 

Raster  cell 

Number  of  line  segments  per  line 

resolution 

1 

2 

3 

4 

5 

5  m 

1.12(3) 

1.91  (6) 

1.64  (7) 

1.95  (7) 

2.15(8) 

10m 

0.57  (2) 

0.93  (3) 

0.84  (3) 

0.98  (3) 

1.07  (4) 

20  m 

0.32(1) 

0.54  (2) 

0.47  (2) 

0.52  (2) 

0.61  (2) 

30  m 

0.16(1) 

0.38(1) 

0.33(1) 

0.38(1) 

0.45(1) 

40  m 

0.12(1) 

0.28(1) 

0.26(1) 

0.29(1) 

0.35(1) 

the  LEI  index  decreases  as  does  the  maximum  number  of  cells  deviation  for  the 
sample  lines.  The  second  relationship  is  that  more  complex  lines  have  higher  LEI 
values  and  maximum  cell  deviation  values,  indicating  that  complex  lines  introduce 
more  error  than  simple  lines.  Multiple  linear  regressions  indicated  that  the  both  these 
relationships  are  significant  at  the  a  =  0.05  level  of  significance.  It  is  also  interesting 
to  note  that  incorporating  the  map  registration  error  into  the  regression  model  did  not 
significantly  improve  the  prediction  of  error  (a=0. 10). 

CONCLUSIONS 

In  the  EPPL7  GIS,  the  summed  standard  deviation  reported  during  map  registration 
appears  to  be  a  poor  indicator  of  locational  accuracy  of  the  resulting  map.  Given  the 
map  scale,  the  point  and  line  representation  equated  to  just  under  13  m,  and  so 


340 


perhaps  a  better  estimation  of  adequate  raster  cell  size  would  be  the  13  m  plus  two 
standard  deviations.  This  would  suggest  a  cell  size  of  20  meters. 

It  appeared  that  multiple  map  registrations  increased  error,  but  the  tests  were 
uncontrolled  and  so  no  comments  about  the  significance  of  any  apparent  numerical 
differences  can  be  made. 

In  line  data,  error  is  strongly  correlated  to  line  complexity.  This  makes  good 
intuitive  sense,  as  the  more  complex  the  line,  the  greater  the  number  of  vertices  are 
required  to  represent  that  line  with  reasonable  accuracy.  The  correlation  between  line 
error  and  target  raster  cell  size  is  an  indication  that  larger  raster  cells  do  a  poor  job 
of  calculating  distance,  and  as  the  raster  cell  size  increases,  differences  between  the 
digitized  line  and  the  "truth"  line  that  are  less  than  the  raster  cell  size  are  rounded  to 
zero.  However,  for  maps  covering  large  areas,  small  raster  cell  size  requires  a  great 
deal  of  processing  time  and  storage. 

This  preliminary  work  has  identified  the  need  to  incorporate  various  map  scales  into 
the  procedure,  particularly  larger  scale  maps  that  will  better  allow  the  testing  of 
registration  error  on  overall  locational  error.  Testing  for  multiple  digitizing  personnel, 
and  various  levels  of  operator  experience  might  also  prove  interesting. 

ACKNOWLEDGMENTS 

Special  thanks  go  to  Linda  Delay,  a  student  worker  that  performed  the  digitizing 
and  ran  the  computer  macros  to  determine  the  error  components.  Without  her  time 
and  effort  on  data  collection,  this  project  would  still  be  in  the  conceptual  stage. 

REFERENCES 

Antenucci,  J.  C,  K.  Brown,  P.  L.  Croswell,  M.  J.  Kevany,  and  H.  Archer.  1991. 

Geographical  information  systems:  A  guide  to  the  technology.  Van  Nostrand 

Reinhold,  New  York,  NY.  301  p. 
Aronoff,  S.  1989.  Geographic  Information  Systems:  A  management  perspective. 

WDL  Publications,  Ottawa,  Canada.  294  p. 
Burrough,  P.  A.  1986.  Principles  of  geographical  information  systems  for  land 

resources  assessment.  Oxford  University  Press,  New  York,  NY.  194  p. 
Chrisman,  N.  R.  1991.  The  error  component  in  spatial  data.  In  Geographical 

Information  Systems,  Volume  1 :  Principles,  D.  J.  Maguire,  M.  F.  Goodchild,  and 

D.  W.  Rhind,  editors.  John  Wiley  and  Sons,  New  York.  Pp.  165-174. 
Fisher,  P.  F.  1991.  Spatial  data  sources  and  data  problems.  In  Geographical 

Information  Systems,  Volume  1 :  Principles,  D.  J.  Maguire,  M.  F.  Goodchild,  and 

D.  W.  Rhind,  editors.  John  Wiley  and  Sons,  New  York.  Pp.  175-189. 
LMIC.  1992.  EPPL7  User's  Guide,  Release  2.0,  Tutorial  Chapter,  page  110.  Land 

Management  Information  Center,  St.  Paul,  MN. 


341 


Maffini,  G,  M.  Arno,  and  W.  Bitterlich.  1989.  Observations  and  comments  on  the 
generation  and  treatment  of  error  in  digital  GIS  data.  In  Accuracy  of  spatial 
databases,  M.  Goodchild  and  S.  Gopal,  editors.  Taylor  and  Francis,  Bristol,  PA. 
Pp.  55-68. 


BIOGRAPHICAL  SKETCH 

Matthew  H.  Pelkki  is  an  Assistant  Professor  of  Forest  Management  and  Economics 
at  The  University  of  Kentucky  Department  of  Forestry.  He  graduated  with  a  B.S.F. 
in  1985  from  the  University  of  Michigan's  School  of  Natural  Resources  & 
Environment,  and  then  earned  an  M.S.  (1988)  and  Ph.D.  (1992)  from  the  University 
of  Minnesota's  College  of  Natural  Resources.  Matthew  has  been  an  assistant 
professor  at  the  University  of  Kentucky  since  1991  where  he  teaches  courses  in  timber 
management,  integrated  forest  resource  management,  and  applications  of  GIS  in 
natural  resources.  His  research  interests  include  dynamic  programming  and  stand- 
level  optimization,  natural  resource  information  system  planning  and  design,  and  the 
economics  of  bioremediation/ecological  restoration. 


342 


GPS  Located  Accuracy  Assessment 
Plots  on  the  Modoc  National  Forest 

Kevin  Casey 1 

Abstract.  A  random  grid  of  field  plots  were  evaluated  in  a  blind  test  to  measure 
the  accuracy  of  remotely  sensed  vegetation  polygons  on  the  Modoc  National 
Forest.  Field  crews  gathered  integrated  inventory  data  on  all  vegetation,  as  well 
as  collecting  accuracy  assessment  data.  GPS  positions  were  collected  at  plot 
centers  and  later  post-processed  to  achieve  2-5  meter  horizontal  accuracy.  Plots 
were  not  rotated  into  condition  and  field  crews  had  no  knowledge  of  the  location 
of  mapped  polygons  or  the  label  attributes  of  the  polygons.  This  methodology 
allows  for  collection  of  accuracy  assessment  data  at  the  same  time  that  inventory 
data  is  collected,  thereby  eliminating  the  need  for  separate  crews  and  multiple 
visits  to  the  field. 


INTRODUCTION 

Forest  vegetation  inventories  are  often  based  upon  a  polygon  map  derived  from 
remote  sensing  tools.  In  Region  5  of  the  U.S.  Forest  Service  (California) 
vegetation  maps  are  produced  at  ten  year  intervals  using  a  combination  of  TM 
imagery,  color  aerial  photography  and  field  training  sites.  The  inventory  plots  are 
located  using  a  randomized  grid  of  points.  The  standard  grid  used  in  the  western 
U.S.A.  has  points  at  5000  meter  intervals.  A  densified  grid  is  used  for  conifer 
forest,  having  points  at  2500  meter  intervals.  The  grid  of  plots  is  used  to  gather 
basic  vegetation  composition,  structure  and  growth  data.  From  these  data  a  forest 
inventory  can  be  compiled.  In  addition,  data  needed  to  perform  a  formal  accuracy 
assessment  of  the  vegetation  map  can  be  collected.  This  methodology  saves  time 
and  expense  by  combining  the  data  collection  needs  for  both  the  inventory  and  the 
accuracy  assessment  into  one  project. 

MAPPING 

The  vegetation  map  for  the  Modoc  National  Forest  is  based  upon  1991  TM 
imagery  which  has  been  terrain  corrected  and  georeferenced.  Polygons  are  created 
using  a  clustering  algorithm  and  labels  are  attached  to  polygons  using  a 
combination  of  supervised  and  unsupervised  classification.  The  minimum  mapping 
unit  is  approximately  one  hectare,  or  ten  pixels.  The  CALVEG  system  of 
vegetation  classification  is  used  to  label  polygons.   In  addition,  for  conifer  and 

1  Project  Leader  for  integrated  resource  inventories,  Region  5,  U.S.  Forest  Service  Remote 
Sensing  Lab,  Sacramento,  California. 


343 


hardwood  types,  a  crown  model  is  used  to  predict  stand  size  class  and  density 
class.  The  mapping  generally  occurs  prior  to  the  inventory  field  work,  but  since 
the  inventory  grid  is  entirely  independent  of  the  map,  field  data  collection  can 
occur  before  or  after  the  map  is  completed. 

PLOT  LOCATION 

Office  Procedures 

The  5000  meter  (or  2500  meter)  grid  is  imported  into  a  GIS  in  the  form  of  a  text 
file  containing  UTM  coordinate  pairs  (MOSS  format).  It  is  clipped  with  the  forest 
administrative  boundary  and  then  viewed  on-screen  with  a  SPOT  panchromatic 
image  as  a  backdrop.  Each  point  of  the  grid  is  manually  transferred  to  a  color 
aerial  photograph.  The  photographs  are  pin-pricked  and  labeled  on  the  back  with 
a  plot  number.  The  plots  are  also  transferred  to  1:24,000  scale  topographic  maps 
to  aid  the  field  crews  in  locating  the  plots.  UTM  coordinates  for  each  plot  are 
printed,  also  to  aid  the  field  crew  with  the  plot  location  on  the  ground.  No 
information  regarding  the  vegetation  polygon  map  is  put  on  any  of  the  materials 
given  to  the  field  crew. 

Field  Procedures 

Field  crews  can  use  any  of  the  supplied  materials  to  locate  the  plot  center  on  the 
ground.  The  pin-pricked  photo  is  the  control,  and  a  tolerance  of  20  meters  is 
allowed  for  error.  Crews  are  equipped  with  a  GPS  receiver,  and  it  can  be  used  to 
navigate  to  the  approximate  plot  center  by  entering  the  UTM  coordinates  of  the 
plots  into  the  receiver.  Final  navigation  must  be  done  using  the  aerial  photograph. 

The  Cluster  Plot 

A  five  point  cluster  plot  is  installed,  centered  on  the  pin-pricked  location.  (See 
Figure  1.)  The  total  size  of  the  cluster  plot  is  approximately  one  hectare.  At  each 
point  of  the  cluster,  vegetation  data  is  measured  and  entered  into  a  data  recorder. 
A  series  of  circular,  nested  fixed-area  plots  are  used  for  collecting  vegetation  data. 
Trees  are  measured  individually;  shrubs,  grasses  and  forbs  are  tallied  by  species 
and  percent  cover.  The  inventory  also  collects  data  on  snags,  down  logs  and 
woody  debris  of  various  sizes. 


344 


FIA5  Point  Cluster  Plot 

1  H/ctare  Circle 

^  Plot  Origin  \ 

(O 

^Hectare  Subplot  / 

Figure  1.  —Plot  Configuration 

The  CALVEG  Accuracy  Assessment  Form 

An  accuracy  assessment  form  was  developed  to  capture  an  independent 
assessment  of  the  vegetative  cover  at  each  sub-plot  of  the  cluster  plot.  (See  Figure 
2.)  The  crew  uses  a  dichotomous  key  to  select  the  Primary  Cover  Type  and  Best 
CALVEG  Type  for  each  of  the  five  subplots.  Cover  Types  are  limited  to  a  rather 
short  list;  examples  being  Conifer,  Hardwood,  Sagebrush  shrub,  Mixed  Chaparral, 
Herbaceous,  etc.  Each  Cover  Type  contains  many  different  CALVEG  Types. 
Some  examples  for  the  Conifer  type  are:  Mixed  Conifer-pine  (MP),  Mixed 
Conifer-fir  (MF),  East-side  Ponderosa  pine  (EP),  Western  juniper  (WJ),  and  so  on. 

The  plots  occasionally  fall  on  an  ecotone,  or  boundary,  between  two  or  more 
distinctive  types.  Condition  codes  are  used  to  capture  this  information.  Condition 
A  is  used  for  the  vegetation  a  point  #1.  If  another  distinctive  type  occurs 
elsewhere  on  the  cluster,  it  becomes  Condition  B.  For  Conifer  and  Hardwood 
types,  the  crew  is  instructed  to  evaluate  the  best  size  class,  and  best  density  class 
for  the  stand  of  trees  occupying  the  plot.  In  affect,  we  ask  the  crew  to  provide  the 
"best  fit"  map  label  for  the  vegetation  at  each  of  the  five  subplots. 


345 


USDA  -  FOREST  SERVICE                                                                                                                  REGION  5 

CALVEG  ACCURACY  ASSESSMENT  RECORD 

COMMENTS: 

PLOT/STAND  IDENTIFICATION 

NATIONAL  FOREST            PU           STRATUM                PLOT  NO. 
DISTRICT              COMPARTMENT                STAND  NO. 

X  POINT  NO.  r- 

X  CONDITION  (A,B,C)  ro 

X  PRIMARY 

X  COVER  « 
X  TYPE 

X  BEST  CALVEG  ^ 
X  TYPE 

X  NEXT  BEST 
X  CALVEG  TYPE 

X  BEST  SIZE  CLASS  P> 

X  SECOND  BEST  SIZE  ^ 

X  BEST  DENSITY  CLASS  00 

X  SECOND  BEST  DENSITY  <° 

X  SECONDARY 

X  COVER  -± 

X  TYPE  P 

X  BEST  CALVEG 

X  TYPE  r'' 

12. 

LU 
CL 
H  >- 
CO  1- 

%o 

LU 

X  3 
LU  < 
Z  O 

X  X 

X  BEST  SIZE  CLASS  G> 

X  SECOND  BEST  SIZE  £ 

X  BEST  DENSITY  CLASS  ai 

X  SECOND  BEST  DENSITY  o> 

Figure  2.  -CALVEG  data  entry  form. 


Fuzzy  Logic  Rules 

Often  the  vegetation  on  a  plot  does  not  fit  exactly  with  the  CALVEG 
descriptions,  or  has  a  mosaic  of  tree  sizes  and  densities  occurring  on  the  plot.  For 
these  situations  "Next  Best"  ratings  can  be  assigned  for  any  of  these  parameters. 
This  data  provides  additional  information  about  trends,  and  allows  for  "fuzzy" 
interpretation  of  the  accuracy  data.  It  is  recognized  that  there  is  value  in 
determining  not  only  if  the  map  label  is  right  or  wrong,  but  how  right  or  how 
wrong.  The  "Next  Best"  data  may  indicate  that  the  map  label  is  wrong,  but 
understandably  wrong  given  that  the  "Next  Best"  answer  matches  the  label. 

The  Plot  Diagram 

Whenever  a  "split  condition"  plot  occurs,  the  crew  prepares  a  Plot  Diagram 
(See  Figure  3.)  which  shows  the  condition  boundary  in  relationship  to  the  entire 
cluster  plot.  A  split  condition  plot  can  contain  two  entirely  different  cover  types, 
such  as  Conifer  and  Sage  Brush  shrub;  or  two  distinctive  CALVEG  types,  such  as 
Lodgepole  pine  and  Red  fir;  or  can  result  from  a  distinctive  change  in  tree  size  or 
density  occurring  within  a  single  type.  An  example  might  be  a  dense,  undisturbed 
Red  fir  stand  bordering  on  a  shelterwood  unit  with  widely  spaced  seed  trees. 


346 


GPS  OPERATIONS 


Field  Procedures 

Crews  collect  GPS  data  at  the  plot  center  (Point  #1)  only.  The  specifications 
require  a  minimum  of  180  positions  be  logged.  The  receiver  is  set  up  to  record  3- 
D  positions  only.  The  elevation  mask  is  set  at  15  degrees,  the  PDOP  mask  is  set  at 
six  and  the  Signal  Level  mask  is  set  at  six.   These  parameters  were  selected  to 


Plot  Diagram      Plot  285 


GPS  file:  R071819A.SSF  at  Point  #  1  Date:7/18/95 


Figure  3.  —Example  of  split  condition  plot 

produce  high  quality  positions  that  can  be  differentially  corrected  to  yield 
horizontal  accuracy  of  2-5  meters.  Occasionally  it  is  not  possible  to  collect  GPS 
positions  at  the  plot  center  due  to  overhanging  trees  or  the  terrain.  Crews  can 
collect  GPS  positions  at  one  of  the  other  four  subplots,  provided  they  note  the 
location  GPS  data  was  collected  on  the  plot  diagram. 

Post  Processing 

The  U.S.  Forest  Service  operates  three  permanent  GPS  base  stations  in  northern 
California.  A  Susanville  base  station  was  used  to  perform  differential  corrections 
for  the  Modoc  project.  It  is  located  50  to  150  miles  from  the  plot  locations  on  the 
Modoc  forest.  Nearly  100%  percent  success  was  achieved  in  correcting  the  rover 


347 


positions.  The  mean  corrected  position  for  each  plot  was  exported  to  MOSS 
format  and  then  brought  into  a  GIS.  The  actual  GPS  position  of  the  plot  replaces 
the  "predicted"  position  stored  in  the  database.  The  difference  between  the 
predicted  location  of  the  plot  center  and  the  GPS  located  plot  center  is  usually  10- 
40  meters  horizontal  distance.  Figure  4  illustrates  an  example  of  how  a  plot  falls 
within  a  vegetation  polygon.  Note  that  the  "predicted"  plot  location  can  result  in  a 
plot  falling  into  one  polygon,  while  the  GPS  position  falls  into  a  different  polygon. 
The  predicted  plot  location  would  suggest  that  Veg  Type  B  would  best  match  the 
vegetation  data  collected  in  the  field.  However,  the  GPS  plot  location  falls 
primarily  into  Veg  Type  A.  Using  the  predicted  plot  location  for  an  accuracy 
assessment  would  yield  incongruous  results. 


VEG  TYPE  C 


Figure  4.  -  Example  of  Predicted  vs.  GPS  plot  location. 
ACCURACY  ASSESSMENT 

Analysis  of  Data 

Accuracy  assessment  field  data  consists  of  completed  forms  as  shown  in  Figure 
5.  Cover  type,  CALVEG  type,  size  and  density  are  compared  to  the  polygon 
attributes  of  the  polygon(s)  that  intersects  with  the  GPS  plot  location.  In  the  case 
of  multiple-condition  plots,  the  vegetation  map  is  examined  for  conformity  with 
the  condition  boundaries  indicated  on  the  Plot  Diagram  (see  Figure  3  .)  An  semi- 
automated  approach  has  been  developed  for  capturing  the  map  label  data  from  the 


348 


vegetation  map.  It  involves  buffering  the  GPS  plot  locations  with  a  given  radius, 
resulting  in  a  polygon  coverage  of  circles  roughly  one  hectare  in  size.  This 
polygon  coverage  is  overlaid  onto  the  vegetation  coverage  producing  a  frequency 
report  listing  all  the  vegetation  map  labels  intersecting  each  buffered  plot.  The 
frequency  report  is  the  best  indicator  of  how  well  the  vegetation  map  compares 
with  the  actual  vegetation  as  described  by  the  field  crew. 


UtOA  ■  FOAfST  «( «VIC£  *f (MOM  I 

CALVE G  ACCURACY  ASSESSMENT  RECORD 


X  POINT  NO.  -* 

2. 

o 
so 
< 

z 
O 

8 

8 

X 

3. 
XXX 

X  WEST  CALVEQ 

X  TYPE  • 

X  NEXT  BEST 
X  CALVEQ  TYPE 

6. 

to 

10 

o 

Ui 
N 

55 

»- 
to 

2 

X 

7. 

Ui 
M 

CO 

t- 
co 

Ui 
OQ 

O 

z 

5 

o 

* 

X 

X  1EST  DENSITY  CLASS  » 

X  SECOND  BEST  DENSITY  • 

X  SECONDARY 

X  COVER  - 

X  TYPE  9 

X  BEST  CALVEQ  - 
X  TYPE  ^ 

12. 

UI 

a. 

CO  1- 

J"  > 
X  -J 
Ui  < 

z  u 
X  X 

X  BEST  SIZE  CLASS            w  | 

X  SECOND  BEST  SIZE  £ 

15. 

CO 
CO 

5 

o 

>- 
»- 

CO 

z 

UI 

Q  > 
*~ 

to 

141 
• 

X 

X  SECOND  BEST  DENSITY  a» 

1 

Z 

P 

hf 

z 

wf 

ft\F 

2 

P 

SS  0 

?5 

CO" 

WF 

P 

es 

BL 

WF 

Z 

N 

S 

8L 

ceW 

WF 

i 

1 

S 

Q 

Figure  5.  -Example  of  completed  CALVEG  data  entry  form. 


CONCLUSIONS 

Spatial  Accuracy 

There  are  several  aspects  of  "accuracy"  that  can  be  defined  when  evaluating 
forest  vegetation  maps.  Spatial  accuracy  -  or  how  well  a  feature  on  the  map 
matches  with  a  feature  on  the  ground  is  a  key  aspect  that  needs  to  be  evaluated.  In 
forest  vegetation,  often  the  best  measure  of  spatial  accuracy  is  how  well  the 
boundary  between  forest  and  non-forest  is  represented  on  the  map.  A  classic 
example  would  be  a  small  meadow  within  a  stand  of  conifers.  A  perfectly 
registered  vegetation  map  would  depict  the  meadow  in  the  same  location  as  a  GPS 
traverse  of  the  meadow.  Or,  as  in  this  case,  the  condition  boundary  described  by 
the  field  crew  would  match  with  a  polygon  boundary  in  the  vegetation  map. 
Because  maps  derived  from  TM  imagery  have  errors  inherent  with  registration,  as 
well  as  other  factors,  it  is  very  important  that  the  accuracy  assessment  plot  be 
located  with  a  minimum  of  error.  Using  corrected  GPS  data  to  properly  locate  the 
plots  in  the  GIS  is  crucial  to  this  effort.  The  less  uncertainty  there  is  in  the  location 
of  the  accuracy  assessment  plot,  the  better  we  can  evaluate  the  accuracy  of  the 
map  itself. 


349 


Map  Label  Accuracy 

The  accuracy  of  map  labels  in  depicting  the  "right"  vegetation  type  is  another 
aspect  to  evaluate.  This  can  be  done  at  several  levels.  The  vegetation  maps 
contain  a  heiarchy  of  information  about  vegetation:  Life  Form/CALVEG/Size  & 
Density.  At  the  most  basic  level,  we  need  to  know  how  well  the  map  represents 
Life  Forms.  Example:  how  many  of  the  Conifer  plots  are  actually  mapped  as 
Conifer? 

At  the  CALVEG  level  we  need  to  know  how  many  of  the  Red  fir  plots  are 
mapped  as  Red  fir?  And  at  the  next  level  we  need  to  know  how  many  of  the 
Large  Tree/High  Density  Red  fir  plots  are  mapped  as  Large  Tree/High  Density? 

The  accuracy  assessment  plots  installed  on  the  Modoc  National  Forest  provide 
the  data  to  answer  these  questions.  They  also  contain  enough  "fuzziness" 
information  to  allow  for  interpretation  of  wrong  map  labels.  If  map  labels  are 
wrong,  we  want  to  know  if  they  are  only  slightly  wrong,  or  completely  wrong. 
For  users  of  the  vegetation  map,  these  issues  can  be  very  important. 

Project  Efficiency 

Combining  the  collection  of  accuracy  assessment  data  with  the  standard 
inventory  plot  has  improved  the  efficiency  of  the  project.  The  accuracy  data  can 
be  collected  quickly  ,  once  on  site.  As  with  most  field  data  collection  projects,  the 
largest  cost  item  is  getting  to  the  plot.  Provided  that  the  plot  work  can  be 
completed  in  one  day  by  a  two  person  crew,  there  is  minimal  additional  expense 
incurred  by  collecting  the  accuracy  data.  Some  additional  training  is  necessary  for 
crews  to  complete  the  accuracy  assessment  form,  but  the  training  can  be  combined 
with  normal  inventory  procedure  training. 

Using  GPS  to  precisely  locate  plot  centers  is  valuable  from  several  standpoints. 
For  the  accuracy  assessment,  it  provides  control;  a  position  accurate  to  only  a  few 
meters,  that  can  be  compared  with  a  map  product.  If  the  map  contains  registration 
errors,  the  GPS  data  will  show  it  conclusively.  Secondly,  having  precise  GPS 
coordinates  of  the  plot  location  allow  for  independent  evaluation  of  other  map 
products  that  describe  life  form,  cover  or  vegetation.  Thirdly,  the  GPS 
coordinates  provide  a  navigational  aid  in  relocating  the  plots  for  future 
remeasurement. 

REFERENCES 

•  Forest  Inventory  and  Analysis  User's  Guide,  1995.  Published  by  USD  A  Forest 
Service,  Region  5  -  Pacific  Southwest. 

•Integration  of  Inventory  and  Field  Data  for  Automated  Fuzzy  Accuracy 
Assessment  of  Large  Scale  Remote- Sensing  Derived  Vegetation  Maps  in  Region  5 
National  Forests,  1996.  Jeff  A.  Milliken  and  Curtis  E.  Woodcock. 


350 


A  Method  for  Measuring  the  Spatial 
Accuracy  of  Coordinates  Collected  Using 
the  Global  Positioning  System 

Thomas  Owens1  and  David  McConville2 


Abstract.-The  study  applied  a  method  to  measure  the  spatial  accuracy  of  point 
data.  The  quantity  of  interest  was  the  difference  between  global  positioning 
system  (GPS)  coordinates  and  surveyed  coordinates.  This  difference  is  a 
bivariate  quantity  and  the  probability  distribution  function  (PDF)  can  be 
described  by  an  ellipse  with  the  center  at  X  and  Y.  The  confidence  ellipse  is  an 
estimate  of  accuracy;  the  sample  mean  is  or  is  not  significantly  different  from  the 
survey  locations  at  a  given  a.  The  tolerance  ellipse  is  an  estimate  of  precision; 
a  given  percentage  of  the  population  sampled  is  enclosed  in  the  tolerance  ellipse 
at  a  given  a.  Thirty-five  locations  were  measured  with  a  GPS  receiver  at 
surveyed  locations  in  the  Mississippi  River  floodplain  near  La  Crosse,  Wisconsin. 
The  average  offset  was  0.18  m  in  the  easting  (X)  direction  and  -1.13  m  in  the 
northing  (Y)  direction.  Hotelling's  one  sample  test  determined  that  H0  (no 
significant  departure  from  the  survey  locations  exists)  was  rejected  at  the  0.05 
level,  which  indicated  a  systematic  error  in  the  sample.  Ninety-five  percent  of 
the  population  sampled  (at  the  0.05  level)  was  contained  in  an  ellipse  centered 
on  (0.18  m,  -1.13  m),  had  a  major  axis  of  7.49  m,  a  minor  axis  of  5.12  m,  and 
an  angle  of  87.74  degrees. 


INTRODUCTION 


The  Environmental  Management  Technical  Center  acquires  and  uses  spatial 
data  for  inventory  and  analysis  of  trends  on  the  Upper  Mississippi  River  in  the 
implementation  of  the  Long  Term  Resource  Monitoring  Program.  Evaluating 
accuracy  and  precision  of  spatial  data  in  geographic  information  systems  (GIS)  is 
important  for  determining  the  appropriate  use  of  these  data.  However,  a  good 
method  has  not  been  documented  for  measuring  the  accuracy  of  location 
measurements  because  of  the  problem's  complexity  and  the  difficulty  in  obtaining 
accurate  location  information  with  traditional  survey  techniques. 


Cartographer,  Environmental  Management  Technical  Center,  Onalaska,  WI. 
Professor.  St.  Mary  's  University.  Winona.  MN 


351 


It  is  conceptually  easier  to  determine  the  spatial  accuracy  of  points  rather  than 
lines  or  polygons.  A  point  may  represent  a  specific  location  such  as  a  survey 
point  or  it  may  represent  an  area  displayed  at  a  small  scale.  If  a  point  represents 
a  real  single  location,  it  is  possible  to  measure  its  displacement  from  its  "true" 
location.  Lines  and  polygons  are  complex  abstract  objects  and  measuring  their 
spatial  accuracy  is  more  complex  (Goodchild  and  Gopal  1989).  For  this  reason, 
this  study  concentrates  on  the  spatial  accuracy  of  points  and  leaves  the  spatial 
accuracy  of  polygons  for  later  study. 

The  advent  of  the  Global  Positioning  System  (GPS)  has  reduced  the  expense 
and  complexity  of  traditional  survey  techniques  for  measuring  the  location  of 
objects  on  the  earth.  The  GPS  enables  nonsurveyors  to  determine  the  location  of 
objects  with  relative  ease,  but  gives  no  information  about  the  data's  accuracy  and 
precision.  This  study  devises  a  straight-forward,  repeatable,  and  statistically 
sound  method  of  estimating  the  horizontal  accuracy  of  GPS  derived  coordinates. 
Two  measures  of  accuracy  are  widely  used  ,  RMS  and  CEP.  Root  Mean  Square 
(RMS)  error  is  the  square  root  of  the  sum  of  the  square  of  the  euclidian  distances 
of  the  error,  divided  by  the  number  of  observations.  Circle  Error  Probable  (CEP), 
is  defined  as  a  circle  that  will  enclose  50  percent  of  the  dot  print.  These 
measures  have  two  disadvantages:  (1)  they  contain  no  information  about  the 
direction  of  the  error,  and  (2)  they  are  only  measures  of  accuracy,  with  no 
information  about  the  variability  (precision)  of  the  measurements. 

Error  and  accuracy  are  conceptually  the  opposite  sides  of  the  same  coin.  Error 
is  the  difference  between  a  quantity's  correct  value  and  its  estimated  value  (Thapa 
and  Bossier  1992).  Accuracy  is  the  nearness  of  quantities  to  their  true  values 
(Bolstad  et  al.  1990).  Errors  may  be  classified  into  three  types:  (1)  gross  errors 
or  blunders,  (2)  systematic  errors,  and  (3)  random  errors.  Gross  errors  are  caused 
by  carelessness  or  mistakes  of  the  observer  using  the  equipment.  Systematic 
errors  are  due  to  environmental  factors,  instrumental  imperfections,  and  human 
limitations;  they  introduce  bias  into  the  estimate.  Even  after  all  gross  and 
systematic  errors  have  been  eliminated,  some  small  random  errors  will  remain. 

Precision  is  the  similarity  of  repeated  measurements  among  themselves.  In  a 
sample  of  coordinates  representing  the  same  position,  precision  is  the  deviation  of 
the  points  from  the  mean  coordinate.  Total  error  for  a  particular  measurement  is 
the  sum  of  the  mean  (systematic)  and  deviation  (random)  errors.  Average  error 
(accuracy)  and  distribution  about  this  mean  (precision)  characterize  the  positional 
uncertainty  of  digital  spatial  data,  which  is  related  to  the  statistical  distribution  of 
the  errors  and  can  be  modeled  by  a  probability  distribution  function  (PDF).  The 
average  difference  between  the  test  and  control  data  estimates  the  accuracy,  while 
the  frequency  distribution  establishes  the  theoretical  PDF  which  contains 
information  on  both  accuracy  and  precision. 


352 


When  a  sample  of  GPS  coordinates  is  taken  at  surveyed  locations,  the  quantity 
of  interest  is  the  difference  in  GPS  coordinates  from  the  surveyed  coordinates. 
The  X  and  Y  difference  between  GPS  coordinate  estimates  and  the  surveyed 
coordinate  is  a  bivariate  quantity.  The  joint  PDF  can  be  described  by  a 
probability  ellipse  (Batschelet  1981)  with  the  center  at  X  and  Y.  An  ellipse  is  an 
appropriate  shape  for  a  joint  PDF  because  it  has  two  dimensions;  it  is  not 
rectangular  because  the  joint  probability  of  points  occurring  in  the  corners  is 
small,  and  it  is  generally  not  circular,  because  X  and  Y  are  not  necessarily  the 
same.  The  ellipse  can  be  described  with  five  statistics:  the  sample  means  of  (1) 
X,  (2)  and  Y,  the  sample  standard  deviations  of  (3)  X,  (4)  and  Y,  (5)  and  the 
correlation  coefficient  between  X  and  Y.  The  quantities  X  and  Y  are  jointly 
distributed,  where  X  and  Y  depend  on  each  other,  but  different  pairs  are 
independent  of  each  other  in  a  sample. 

There  are  three  ellipses  of  interest:  the  standard  ellipse,  the  confidence  ellipse, 
and  the  tolerance  ellipse.  The  standard  ellipse  is  a  descriptive  tool  used  to 
visualize  the  shape  of  the  ellipse  and  its  orientation.  It  is  readily  calculated  and 
contains  about  40%  of  the  sample  population.  It  is  not  dependent  on  the  sample 
size  and  cannot  be  used  for  statistical  inference.  Five  statistics  are  required  to 
construct  the  ellipse:  X,  Y,  s2  ,  s2  ,  and  the  correlation  coefficient  r  (r  = 
Cov(X,Y)/sxsy  and  Cov(X,Y)  =  l/fa-l)!^  -  X)(Yj  -  Y)),  which  measures  the  joint 
behaviors  of  X  and  Y.  The  center  of  the  ellipse  is  (X,  Y)  and  the  equation  of  the 
ellipse  is  as  follows: 


An  ellipse  is  defined  by  a  major  axis  (a),  a  minor  axis  (b),  and  an  angle  0  of 
translation,  which  is  the  angle  that  the  major  axis  is  offset  from  the  X  axis  (and 
the  angle  the  minor  axis  is  offset  from  the  Y  axis). 

Another  coefficient  must  be  calculated  to  arrive  at  the  ellipse  parameters: 


sy(X  -  X)2  -  2rsxsy(X  -  X)(Y  -  Y)  +  s2(Y  -  Y)  =  (1  -  r^sjs2 


(1) 


For  ease  of  calculation,  some  coefficients  are  shortened: 


(2,  3) 
(4,  5) 


R  =  [(A  -  C)2  +  4B2] 


(6) 


Then  we  can  calculate  the  ellipse  parameters: 


a  =  [2D/(A  +  C  -  R)] 
b  =  [2D/(A  +  C  +  R)] 


1/2 


(7) 
(8) 


1/2 


353 


6  =  arctan[2B/(A  -  C  -  R)] 


(9) 


The  other  two  ellipses  have  an  identical  shape  and  orientation  but  have  different 
major  and  minor  axes.  The  Hotelling's  confidence  ellipse  covers  the  sample's 
center  with  a  given  accuracy  and  estimates  accuracy.  In  addition,  Hotelling's  one- 
sample  test  can  be  used  to  determine  if  the  bivariate  sample  mean  is  significantly 
different  than  0.  The  confidence  ellipse  is  a  region  that  contains  the  population 
mean  at  a  given  probability  with  the  center  of  the  ellipse  at  (X,  Y).  The  equation 
of  the  ellipse  is  as  follows: 

s2y(X  -  X)2  -  2rsxsy(X  -  X)(Y  -  Y)  +  s2(Y  -  Y)  =  (1  -  r^s's2  nlT2  (10) 

Note  that  the  right  side  of  the  expression  has  an  added  term,  T2,  which  is  based 
on  the  familiar  F  value  for  the  univariate  solution  where: 

T2  =  2[(-  l)/(n  -  2)]F2  n_2  (11) 

Fa/i,n-2  denotes  the  critical  value  from  the  F  one-tailed  distribution  with  n-2 
degrees  of  freedom  and  significance  level  of  a.  The  confidence  ellipse  is  then 
calculated  using  the  same  coefficients  described  for  the  standard  ellipse. 

The  tolerance  ellipse  is  an  estimate  of  precision.  It  is  an  estimate  of  the 
confidence  (e.g.,  95%)  of  a  percentage  of  the  population  sampled  (e.g,  95%)  that 
is  enclosed  in  the  tolerance  ellipse.  The  tolerance  ellipse  is  a  region  that  contains 
a  given  percent  of  the  population  at  a  given  probability  with  the  center  of  the 
ellipse  at  (X,  Y).  The  equation  of  the  ellipse  is: 

s2(X  -  X)2  -  2rsxsy(X  -  X)(Y  -  Y)  +  s2(Y  -  Y)  =  (1  -  r^s2  H  (12) 

The  right  side  of  the  expression  has  the  term  H,  which  is  based  on  the  non- 
central  %2  distribution  and  is  approximated  by  Table  2  in  Chew  (1966).  The 
tolerance  ellipse  is  then  calculated  using  the  same  coefficients  described  for  the 
standard  ellipse. 

This  elliptical  method  of  describing  spatial  accuracy  and  precision  is  valid 
because  it  is  (1)  intuitive,  (2)  understandable  graphically,  (3)  statistically  sound, 
(4)  sufficient  (e.g.,  it  describes  all  relevant  parameters),  and  (5)  concise. 

METHODS 

The  location  of  the  study  included  Navigation  Pools  7  and  8  of  the  Upper 
Mississippi  River  near  La  Crosse,  Wisconsin.  This  reach  of  the  river  is  a  complex 
of  islands,  channels,  and  backwaters  surrounded  by  500-ft  high  bluffs. 


354 


Forty-one  surveyed  ground  locations  were  used  as  test  points.  Thirty-six  third- 
order  points  were  surveyed  by  the  U.S.  Army  Corps  of  Engineers  and  were 
distributed  along  the  main  channel  of  the  river.  Five  first-order  points  located  at 
road  intersections  in  the  floodplain  were  surveyed  by  the  Wisconsin  and 
Minnesota  Department  of  Highways. 

A  Trimble  Pathfinder  Basis  Plus  GPS  receiver  was  used  in  the  field  as  the 
rover.  The  GPS  measurements  were  taken  with  the  rover  on  June  8,  9,  and  17. 
1994.  The  PDOP  mask  was  set  to  5.  The  receiver  was  set  to  take  one  position 
per  second  and  the  horizon  mask  was  set  at  10°.  The  unit  received  signals  until 
approximately  200  readings  were  taken  at  each  point.  A  Trimble  Pathfinder 
Professional  receiver  was  established  at  a  surveyed  point  to  act  as  the  base  station. 
The  base  was  set  to  take  positions  once  every  5  seconds,  had  a  PDOP  mask  of  5, 
and  a  horizon  mask  of  10°.  The  measurements  were  post-processed  to  introduce 
differential  correction  using  Trimble's  PFinder  software  on  a  personal  computer. 

Survey  measurements  were  recorded  in  the  State  Plane  Coordinate  System, 
Minnesota  South  zone,  and  converted  to  the  Universal  Transverse  Mercator 
(UTM)  projection,  zone  15.  The  GPS  measurements  were  taken  in  WGS-84  and 
converted  to  the  UTM  projection,  zone  15,  NAD27. 

The  X  (easting)  and  Y  (northing)  differences  between  the  survey  GPS 
measurements  were  tested  to  determine  if  their  distribution  was  normal  using  the 
X2  (chi-square)  test  (Zar  1984).  If  the  easting  and  northing  differences  were 
normally  distributed,  it  is  possible  to  use  bivariate  normal  tests  to  test  for 
systematic  bias  from  zero  and  to  describe  the  PDF  with  ellipses  (Batschelet  1981). 
Hotelling's  confidence  ellipse,  the  standard  ellipse,  and  the  tolerance  ellipse  (Chew 
1966)  were  calculated  to  described  the  PDF.  Hotelling's  one-sample  test  was  run 
to  determine  if  the  sample  average  was  significantly  different  than  0. 

RESULTS  AND  DISCUSSION 

Thirty-five  usable  locations  were  differentially  corrected  from  the  41  locations 
measured.  Five  of  the  original  locations  were  not  usable  because  the  base  station 
and  rover  did  not  receive  positions  from  the  same  satellites  and  so  the  GPS 
readings  were  not  corrected  during  differential  post-processing  (possibly  due  to 
setting  the  mask  angle  to  10°  on  both  the  rover  and  the  base  station).  One 
surveyed  point  had  incorrect  coordinates  and  was  discarded. 

The  average  distance  from  the  survey  points  was  -1.13  m  in  the  northing  (Y) 
direction  and  0. 15  m  in  the  easting  (X)  direction. 


355 


The  easting  and  northing  distances  were  tested  for  normality.  H0  (that  the 
distributions  were  normal)  was  not  rejected  at  the  0.05  level.  Thus,  we  could 
calculate  the  descriptive  and  statistical  ellipses  and  test  to  determine  if  there  was 
a  significant  difference  from  0  for  the  mean  distance. 

The  standard  ellipse  was  centered  on  the  sample  means  of  (0.15,  -1.13),  and  had 
a  major  axis  of  2.60  m  and  a  minor  axis  of  1.77  m  with  an  angle  from  the  X  axis 
of  87.74°  (figure  1).  The  confidence  ellipse  had  a  major  axis  of  1.12  m  and  a 
minor  axis  of  0.77  m,  with  the  same  angle  of  87.74°.  Hotelling's  one-sample  test 
was  used  to  determine  if  the  sample  means  were  significantly  different  than  0. 
The  result  was  that  H0  (that  the  population  means  equal  0)  was  rejected  at  the  0.05 
level.  The  tolerance  ellipse,  which  was  calculated  to  contain  95%  of  the 
population  at  the  95%  confidence  level,  had  a  major  axis  of  7.49  m  and  a  minor 
axis  of  5.12  m. 

The  confidence  ellipse  does  not  encompass  the  origin  (0,  0),  which  graphically 
shows  that  the  sample  mean  is  different  than  0.  Thus,  the  sample  has  systematic 
error  that  is  offset  in  the  south  and  east  directions.  We  can  also  say  that  we  are 
95%  confident  that  95%  of  the  population  we  sampled  is  contained  in  an  ellipse 
that  is  centered  on  0.15,  -1.13,  and  has  a  major  axis  of  7.49  m,  a  minor  axis  of 
5.12  m  with  an  angle  of  87.74°  at  the  0.05  level. 

CONCLUSION 

The  purpose  of  this  study  was  to  apply  a  method  to  measure  the  spatial 
accuracy  of  point  data.  Hotelling's  one-sample  test  made  inferences  about  the 
sample  data's  means  and  a  tolerance  ellipse  described  the  PDF  of  the  population. 
The  confidence  ellipse  is  an  intuitive  graphic  that  shows  whether  systemic  error 
exists  in  the  sample.  This  method  is  theoretically  sound  and  provides  intuitive  and 
concise  statistics  about  a  population.  We  can  now  describe  the  accuracy  of  the 
sample  data  generated  by  GPS  described  in  the  study  as  having  a  mean  that  is 
0.18,  -1.13  m,  and  has  systemic  error  in  the  east  and  south  directions.  We  can 
further  state  that  95%  of  the  population  is  contained  in  an  ellipse  (at  the  95% 
confidence  level)  with  a  center  at  0.18,  -1.13  m  having  a  major  axis  of  7.49  m, 
a  minor  axis  of  5.12  m,  and  an  angle  of  translation  from  the  horizontal  of  87.74°. 

This  method  to  measure  and  describe  the  spatial  accuracy  of  point  data  is  useful 
for  three  reasons.  First,  the  method  is  repeatable;  similar  results  will  be  obtained 
for  a  given  set  of  points  when  the  method  is  repeated  and  accuracy  statements  can 
be  compared  for  different  sets  of  data.  Second,  it  is  statistically  valid;  this  method 
provides  a  probability  statement  about  the  accuracy  statement.  Third,  it  is 
sufficient.  Merely  stating  that  the  GPS-derived  points  are  "within  5  m  of  their 


356 


true  location"  is  not  sufficient  because  no  probability  statement  is  associated  with 
the  accuracy  statement.  In  addition,  X  and  Y  coordinates  are  related,  but  do  not 
vary  in  a  perfect  one-to-one  relationship;  it  is  necessary  to  describe  how  both  X 
and  Y  vary,  along  with  the  confidence  ellipse,  which  describes  the  accuracy,  and 
the  tolerance  ellipse,  which  describes  the  precision. 


15 


10 


<u 

.03 


1|  0 

o 


00 

o 


-5 


•10 


-15 


-15 


GPS  Offsets  from  Survey  Points 

Comparison  of  Survey  Points  and  GPS  Points 


-5  0 
Meters 


GPS  Easting  Offset 


 ( 

□  ^ 



/ 

/ 

/ 

/  \ 

/  \ 

\ 

.  1 

 1 — 1 — /ta 

B 

EE 

K/ 

\ 

%  \  \ 

pa  V  \ 

*  / 
/ 

E 

\ 

I  \ 

Confidence 
'  Standard  E 

j  Ellipse 
llipse 

1  Tolerance 

Ellipse 

15 


Figure  1. -Distribution  of  distances  between  the  GPS-derived  cooidinates  and  survey-derived 
coordinates  and  ellipses  which  describe  the  sample,  accuracy,  and  probability  distribution  function. 


357 


ACKNOWLEDGEMENTS 


This  study  was  supported  by  the  U.S.  Army  Corps  of  Engineers  and  the 
National  Biological  Service  through  the  Upper  Mississippi  River  System  Long 
Term  Resources  Monitoring  Program. 

REFERENCES 

Batschelet,  E.   1981.  Circular  statistics  in  biology.  Academic  Press,  New  York. 
310  pp. 

Bolstad,  P.,  P  Gessler,  and  T.  Lillesand.  1990.  Positional  uncertainty  in  manually 

digitized  map  data.  International  Journal  of  GIS.  4(4):  399-412. 
Chew,  V.  1966.  Confidence,  predictive,  and  tolerance  regions  for  the  multivariate 

normal  distribution.  American  Statistical  Association  Journal  61:605-617. 
Goodchild,  M.,  and  S.  Gopal.   1989.  Preface.  Pages  xi-xv  in  National  Center  for 

Geographic  Information  and  Analysis.    M.  Goodchild  and  S.  Gopal,  editors. 

Accuracy  of  Spatial  Database.  Taylor  and  Francis,  New  York. 
Thapa,  K.,  and  J.  Bossier.     1992.     Accuracy  of  spatial  data  used  in  GIS. 

Photogrammetric  Engineering  and  Remote  Sensing  58(6): 835-841. 
Zar,  J.  1984.  Biostatistical  analysis.  Prentice-Hall,  Englewood  Cliffs,  New  Jersey. 

620  pp. 

BIOGRAPHICAL  SKETCHES 

Thomas  Owens  is  a  Cartographer  with  the  National  Biological  Service's 
Environmental  Management  Technical  Center  in  Onalaska,  Wisconsin.  He 
received  a  Master  of  Science  in  forest  management  and  remote  sensing  from  the 
University  of  Minnesota  in  1978.  He  is  Leader  of  the  Cartographic  Services 
Group  and  provides  GIS  expertise  to  the  EMTC. 

David  McConville  is  a  codirector  of  the  Resource  Studies  Center  and  Professor 
of  Biology  at  Saint  Mary's  University  of  Minnesota.  He  holds  a  Ph.  D.  in 
fisheries  from  the  University  of  Minnesota  and  has  over  25  years  of  experience 
working  on  aquatic  ecosystems  and,  in  particular,  Mississippi  River  ecology. 


358 


GPS  vs  Traditional  Methods  of  Data 
Accuracy  Input:  Improving  Spatial  Data 

Accuracy? 

Russell  Combs,  Jr.  l,  James  L.  Smith2,  and  Paul  V.  Bolstad3 


Abstract. -Geographic  information  systems  (GIS)  have  become  an  integral 
part  of  many  natural  resource  organizations.  The  increased  interest  and 
use  of  these  systems  for  a  variety  of  managerial  tasks  has  resulted  in  an 
increased  concern  about  the  accuracy  of  spatial  databases.  Many  natural 
resource  organizations  are  using  their  GIS  to  perform  complex  analysis 
and  they  are  becoming  very  aware  of  what  the  cost  might  be  for  using 
inaccurate  spatial  data.  Thus  many  organizations  are  involved  in 
developing  procedures  or  methods  for  improving  and  maintaining  the 
quality  and  accuracy  of  their  spatial  data.  This  paper  will  focus  on  the 
issue  of  spatial  data  accuracy  and  how  it  can  be  improved  or  maintained 
by  natural  resource  organizations.  In  particular,  the  paper  will 
concentrate  on  one  of  the  main  areas  influencing  spatial  data  accuracy, 
data  input.  The  paper  will  review  and  compare  the  positional  accuracy 
levels  of  some  of  the  most  common  methods  of  data  input  (manual 
digitizing)  to  newer  technology  and  techniques  (GPS,  PC-based  single- 
photo  space  resection)  now  available  to  natural  resource  organizations. 


INTRODUCTION 

Positional  accuracy  in  a  natural  resource  GIS  can  be  defined  as  a  measure, 
usually  in  ground  distance,  of  how  well  the  digital  object  coordinates  in  the  various 
spatial  data  layers  correspond  to  the  "true"  coordinates  of  an  entity  on  the  ground 
(Bolstad  &  Smith  1992,  Chrisman  1991,  and  Blakemore,  1984).  The  accuracy  of 
a  spatial  database  is  a  direct  function  of  the  methods  used  to  input  the  spatial  data, 
and  the  sources  of  data  used  to  collect  the  spatial  information.  As  with  any 
mapping  procedure  , whether  digital  or  cartographic,  generalizations  are  made  and 
database  coordinates  used  to  represent  entities  may  differ  from  their  "true"  ground 
locations.  Thus,  the  accuracy  of  a  spatial  database  is  inherently  limited  by  the 
method  or  sources  of  data  used  to  construct  the  database. 

A  natural  resource  GIS  spatial  database  usually  includes  multiple  data  layers,  with 
each  layer  representing  a  different  theme  or  group  of  similar  features. 

123 


1  Database/GPS  Forester.  Canal  Forest  Resources,  Inc.,  Charlotte,  NC 

2  Manager  GIS  and  Remote  Sensing  Services.  Canal  Forest  Resources,  Inc..  Charlotte,  NC 

3  Assistant  Professor.  Department  of  Forest  Resources,  University  of  Minnesota 


359 


The  individual  data  layers  in  a  natural  resource  GIS  are  not  typically  developed  or 
created  using  the  same  methods  or  sources  of  spatial  data.  Natural  resource 
organizations  routinely  deal  with  ground  features  that  are  not  well  mapped  or  must 
work  with  with  poorly  or  undocumented  maps.  Thus,  a  wide  range  of  data  sources 
are  often  used,  and  often  two  different  sources  of  data  are  used  within  a  given  data 
layer.  This  leads  to  a  combination  of  methods  and  sources  being  used  to  develop 
or  update  the  spatial  database,  and  in  turn  leads  to  a  spatial  database  with  many 
different  data  layers  having  varying  levels  of  positional  accuracy.  This  paper  will 
review  the  accuracies  of  the  most  common  data  input  methods/sources,  and 
propose  some  suggestions  on  how  natural  resource  organizations  can  improve  the 
positional  accuracy  of  their  spatial  databases.  In  addition,  the  paper  will  discuss 
what  if,  any  accuracy  or  benefit  may  be  gained  by  using  some  of  the  new 
technology  becoming  available. 

Traditional  Methods/Sources  of  Spatial  Data 

The  traditional  methods  and  sources  of  spatial  data  input  for  natural  resource 
organizations  can  be  categorized  into  two  main  types.  1)  manually  digitizing  from 
large-scale  maps  or  aerial  photographs,  and  2)  the  use  of  purchased  digital  data 
(DLG,  TIGER,  DEM).  While  purchased  digital  data  is  becoming  more 
widespread  throughout  the  natural  resource  community,  its  use  is  limited  to  those 
organizations  with  established  GIS  systems,  and  those  who  have  familiarity  with 
these  type  of  data.  Manual  digitization  is  still  by  far  the  most  common  method  of 
data  input,  and  thus  the  major  source  of  positional  error  in  spatial  databases. 
Positional  errors  resultant  from  manual  digitization  are  generally  a  function  of 
three  sources:  media  limitation,  user  error,  and  control  point  or  coordinate 
registration  error  (Walsh  et  al,  1987). 

While  all  three  of  these  factors  vary  as  to  the  affect  they  have  on  the  positional 
accuracy  of  manually  digitized  data,  the  first  of  these  factors,  source  media 
limitation,  has  the  largest  impact  on  the  amount  of  positional  error  introduced  to 
spatial  data  during  manual  digitization.  Manually  digitized  data  in  natural  resource 
organizations  are  usually  collected  from  either  aerial  photography  or  large-scale 
maps.  In  both  cases  the  source  data  has  inherent  limitations  that  affect  the 
accuracy  of  the  data.  Aerial  photography  is  affected  by  tilt  and  terrain 
displacement,  while  many  large-scale  natural  resource  maps  contain  significant 
positional  errors  due  to  poor  drafting,  incorrect  boundary  placement,  or  drafting 
from  aerial  photography. 

The  amount  of  positional  error  that  can  be  introduced  to  spatial  data  because  of 
these  factors  has  been  documented,  and  by  reviewing  previous  research,  an 
estimate  of  the  positional  error  caused  by  manual  digitization  can  be  obtained.  For 
example,  horizontal  positional  errors  of  up  to  73.0  meters  in  steep  terrain  and 


360 


15.0  meters  in  low  terrain  have  been  observed  for  spatial  data  digitized  directly  off 
of  aerial  photography  (Bolstad,  1992).  On  the  other  hand,  large  scale  maps  often 
have  linear  or  areal  features  that  have  line  widths  which  vary  between  0.25  -1.01 
mm  (Bolstad  &  Smith,  1992),  corresponding  to  widths  on  the  ground  of  anywhere 
between  6.0  -  24.0  meters  if  the  data  were  collected  from  a  1 :24,000  scale  map. 
Another  study  observed  line  locations  to  vary  by  0. 127  mm  of  their  true  positions 
on  a  1 :24,000  scale  map,  equivalent  to  3.04  meters  in  actual  ground  distance 
(Dunn  et  al,  1990).  As  can  be  seen,  it  does  not  take  much  variation  in  a  source 
document  to  start  adding  up  to  large  amounts  of  inaccuracy  on  the  ground.  A 
factor  that  is  sometimes  overlooked  when  dealing  with  linear  and  positional 
accuracies,  is  the  associated  area  accuracy  that  is  directly  related  to  the  linear 
accuracy  of  the  spatial  database.  For  instance  Chen  and  Finn  (1994)  stated  that 
average  manually  digitized  area  error  is  low,  with  manually  digitized  data  generally 
underestimating  polygon  area  by  1-3%.  In  addition,  Wiles  (1988)  observed 
acreage  errors  of  up  to  10.23%  for  data  digitized  directly  from  aerial  photography 
in  varying  terrain. 

The  second  source  of  positional  error  resultant  from  manually  digitization  is 
from  operator  or  user  error.  Operator  error  is  the  error  resultant  from  the 
operator's  inability  to  trace  linear  or  point  features  with  the  digitizing  puck. 
Operator  digitizing  ground  errors  of  up  to  1 . 1 5  m  in  the  x-direction  and  1 . 1 7  m  in 
the  y-direction  have  been  documented  for  manually  digitized  data  from  1 : 24,000 
scale  maps  (Combs  1995,  Warner  &  Carson,  1991).   In  addition,  operator 
inaccuracies  of  0.95  m  in  the  x-direction  and  1.26  m  in  the  y-direction  have  been 
calculated  for  manually  digitized  data  collected  from  NAPP  photography  (Combs, 
1995). 

In  addition  to  source  errors  and  user  errors,  positional  accuracy  during  manual 
digitizing  is  also  affected  by  control  point  or  coordinate  registration.  This  is  an 
issue  that  is  often  overlooked  by  natural  resource  organizations,  but  poor  control 
point  registration  does  affect  the  overall  accuracy  of  the  digitized  data.  A  few 
studies  have  highlighted  the  importance  of  using  accurate  control  points. 
Fernandez  et  al.  (1991)  documented  map  derived  control  point  errors  ranging  from 
2.13  to  7.62  meters  in  ground  distance.  Bolstad  et  al  (1990)  observed  positional 
error  attributed  to  poor  control  points  ranging  from  1.52  to  85.04  meters  on  the 
ground  for  data  digitized  directly  from  aerial  photography  in  varying  terrain. 
Control  point  registration  and  accuracy  can  be  increased  by  using  well  known 
established  point  locations  that  have  been  surveyed  or  through  the  use  of  global 
positioning  systems  (GPS).  Thus,  if  the  positional  errors  associated  with  manual 
digitization  are  additive,  which  they  most  likely  are  to  some  extent,  spatial  data 
collected  using  these  methods  could  be  tens  to  hundreds  of  meters  off  in  positional 
accuracy. 


361 


Non-traditional  Methods/Sources  of  Spatial  Data 


While  manual  digitization  and  the  use  of  purchased  digital  data  are  still  popular 
methods  and  sources  of  spatial  data,  the  development  of  new  data  input  methods 
and  techniques  may  provide  natural  resource  organizations  with  more  accurate  and 
cost-effective  alternatives.  Two  of  these  new  techniques  or  methods  will  be 
discussed:  1)  GPS  and  2)  PC-based  single-photo  space  resection.  While  both  of 
these  alternatives  are  not  necessarily  "new",  they  have  not  been  extensively  applied 
in  natural  resource  organizations. 

GPS  is  a  satellite-based  positioning  system  which  operates  using  L-band  radio 
signals  to  provide  highly  accurate  position,  velocity,  and  time  data.  GPS  can 
provide  accuracy  levels  of  up  to  30.0  cm  (P-code)  and  1-5  meters 
(Course/ Acquisition  Code  or  C/A  code).  Civilian  users  can  achieve  accuracies 
equal  to  those  of  the  P-code  but  generally  this  accuracy  is  reserved  for  military  use 
or  surveyors.  Single-photo  space  resection  is  accomplished  by  measuring  photo- 
coordinates,  via  a  digitizing  tablet,  and  applying  the  collinearity  equations  to  these 
data  to  calculate  ground  coordinates.  Since  the  procedure  is  based  upon  the 
photogrammetric  principles  of  space  resection,  the  calculated  ground  coordinate 
accuracy  is  improved  due  to  the  reduction  of  the  tilt  and  terrain  effects  inherent  in 
the  photograph. 

Recently  studies  have  been  conducted  to  determine  the  accuracy  of  GPS  and 
single-photo  space  resection  in  a  natural  resource  setting.  Evans  et  al.  (1992) 
reported  average  positional  accuracy  of  2.01  m  between  GPS  positions  and  actual 
forest  plot  center  positions  during  navigation  trails  to  locate  known  forest  plot 
centers.  In  other  studies,  applying  differential  correction  to  the  mean  of  100  to 
300  position  fixes  resulted  in  positional  ground  accuracies  of  between  2.0  to  3.0  m 
(August  et  al,  1994)  and  2.0  to  4.0  m  (Deckert,  1994).  A  previous  study  reported 
positional  ground  errors  from  single-photo  space  resections  of  two  different 
mountainous  study  areas  to  be  1 1.39  m  and  7.07  m  respectively  (Ran,  1992).  That 
would  translate  to  about  60.0  m  more  accurate  than  the  data  collected  directly 
from  aerial  photography  in  steep  terrain  as  reported  in  the  studies  previously 
discussed. 

A  more  recent  study  involved  comparing  the  accuracies  of  point  features 
manually  digitized  from  topographic  maps  and  NAPP  photography  versus  the 
accuracies  of  the  same  features  collected  using  single-photo  space  resection  and 
GPS  techniques  (Combs,  1995).  For  this  study,  C/A  code  level  GPS  was  used  to 
determine  the  "true"  ground  or  reference  location  of  the  test  features.  While  the 
GPS  data  collected  for  this  study  can't  determine  the  actual  ground  locations  of  the 
test  features,  previous  work  has  determined  C/A  code  level  GPS  points  to  be 
within  1.0  -  3.0  meters  of  the  "true"  ground  location  (Deckert,  1994).  For  each 


362 


test  point,  300  position  fixes  were  collected  and  averaged  after  differential 
correction  to  determine  a  mean  positional  location  for  the  test  point.  The  manually 
digitized  and  single-photo  space  resection  data  were  then  compared  to  the  GPS 
value  and  the  difference  was  determined  to  be  the  amount  of  positional  error 
attributed  to  that  method/source.  Table  1  highlights  the  results  of  this  study  and 
shows  the  average  positional  accuracy  that  was  determined  by  using  the  various 
techniques  and  sources. 


Table  1.  Mean  positional  error  (meters)  for  two  study  sites  in  Virginia,  GPS  used  as 


"true"  "round  location. 


Data  Set 

High-Relief  Study  Site 

Low-Relief  Study  Site 

Space  Resection 

6.27 

5.64 

Digitized  Aerial  Photograph 

25.86 

7.14 

Topographic  Map 

11.20 

11.64 

The  results  in  Table  1  show  how  spatial  accuracy  can  be  improved  by  using 
either  GPS  or  the  space  resection  technique.  The  positional  accuracy  of  data 
digitized  directly  from  aerial  photography  is  about  what  would  be  expected.  As 
shown  in  Table  1,  the  high-relief  study  site  had  a  mean  positional  accuracy  of 
more  than  3  times  greater  than  the  low-relief  study  site.  This  difference  is 
attributed  to  the  greater  amount  of  terrain  relief  and  tilt  displacement  that  is 
present  in  the  high-relief  study  site  aerial  photography.  In  contrast,  the  single- 
photo  space  resection  data  sets  provided  mean  positional  accuracies  that  are 
almost  equal  between  the  two  study  sites.  Since  the  space  resection  reduced  the 
tilt  and  terrain  effects  inherent  in  the  aerial  photography  (and  assuming  all  other 
factors  are  constant),  the  resultant  mean  positional  accuracies  were  similar  and 
improved  over  just  digitizing  directly  from  the  aerial  photography. 

The  digitized  topographic  map  data  from  both  study  sites  provided  similar 
positional  accuracies  that  were  greater  than  the  NMAS  (National  Map  Accuracy 
Standards),  but  both  were  still  more  than  half  of  what  was  obtained  by  the  space 
resection  technique.  Other  results  that  are  not  directly  shown  in  Table  1  are  the 
accuracy  of  the  GPS  data.  While  GPS  was  only  used  for  the  "true"  ground 
reference  value  for  this  study,  the  results  in  Table  1  show  how  much  each  of  the 
other  data  collection  techniques  differed  in  positional  accuracy  from  the  GPS 
values.  For  example,  on  the  high-relief  study  site  the  GPS  data  was  on  average 
25.0  meters  more  accurate  than  digitizing  directly  from  aerial  photography.  In 
addition,  the  GPS  data  proved  to  be  on  average  11.0  meters  more  accurate  than 
digitizing  directly  from  1 :24,000  scale  topographic  maps.  While  the  GPS  proved 
to  provide  the  greatest  positional  accuracy,  both  it  and  the  single-photo  space 
resection  increased  the  reliability  and  confidence  of  the  data  collected  in  this  study. 


363 


CONCLUSIONS 


The  accuracy  of  digital  spatial  data  collected  by  natural  resource  organizations  is 
affected  by  the  source  of  the  data  and  the  method  use  to  collect  the  data.  As  the 
use  of  GIS  continues  to  grow  in  natural  resource  organizations,  the  reliability  and 
accuracy  of  the  data  used  in  GIS  analysis  is  becoming  increasingly  important. 
Greater  than  any  other  time  in  recent  history,  natural  resource  organizations  are 
being  held  responsible  for  their  actions  and  the  decisions  they  make  concerning  the 
environment.  The  research  and  results  reviewed  in  this  paper  show  that  spatial 
data  collected  using  the  traditional  methods  may  adversely  affect  the  accuracy  of 
spatial  databases.  Research  presented  in  this  paper  showed  that  data  digitized 
directly  from  large-scale  maps  and  aerial  photography  can  provide  positional 
accuracies  anywhere  between  6.0  to  73.0  meters.  On  the  other  hand,  spatial  data 
collected  using  a  PC -based  single-photo  space  resection  technique  can  provide 
positional  accuracies  between  5.0  to  7.0  meters,  depending  on  the  amount  of 
terrain  relief.  In  addition,  GPS  collected  spatial  data  has  been  shown  to  be  capable 
of  providing  positional  accuracies  from  1  to  4  times  more  accurate  than  some  of 
the  traditional  techniques  discussed  in  this  paper.  As  the  use  of  GIS  technology 
continues  to  expand  in  natural  resource  organizations,  a  PC-based  single-photo 
resection  or  more  particular  the  use  of  GPS,  offers  an  alternative  to  traditional 
methods  of  data  entry  and  can  provide  reliable  and  more  accurate  spatial  data. 

REFERENCES 

August,  P.  ,  J.  Michaud,  C.  Labash,  and  C.  Smith.  1994.  GPS  for  Environmental 
Applications:  Accuracy  and  Precision  of  Locational  Data.  Photogrammetric 
Engineering  &  Remote  Sensing.  60(l):41-45. 

Blakemore,  M.  1984.  Generalization  of  Error  in  Spatial  Databases.  Cartographica. 
21:131-139. 

Bolstad,  P.V.  1992.  Geometric  Errors  in  Natural  Resource  GIS  Data:  Tilt  and 

Terrain  Effects  in  Aerial  Photographs.  Forest  Science.  38(2):367-380. 
Bolstad,  P.  V.,  P.  Gessler,  and  T.M.  Lillesand.  1990.  A  Variance  Components 

Analysis  of  Manually  Digitized  Map  Data.  Proc.  ACSM-ASPRS.  Vol  3.  9-18pp. 
Bolstad,  P. V.,  and  J.L.  Smith.  1992.  Errors  in  GIS:  Assessing  Spatial  Data 

Accuracy.  Journal  of  Forestry.  90(1 1):21-29. 
Chen,  Z.  Z.,  and  J.T.  Finn.  1994.  The  Estimation  of  Digitizing  Error  and  Its 

Propagation  with  Possible  Application  to  Habitat  Mapping.  Proc.  International 

Symposium  on  the  Spatial  Accuracy  of  Natural  Resource  Databases.  57-66  pp. 
Chrisman,  N.R.  1991.  The  Error  Component  of  Spatial  Data.  Geographical 

Information  Systems,  Volume  1  Principles.  John  Wiley  &  Sons.  165-173  pp. 
Combs,  R.G.,  1995.  Positional  Accuracy  in  a  Natural  Resource  Database: 

Comparison  of  a  Single-Photo  Resection  Versus  Affine  Registration.  Master's 

Thesis.  VPI&SU.  148  pp. 


364 


Deckert,  C.J.  1994.  Canopy,  Terrain,  and  Distance  Effects  on  Global  Positioning 

System  Position  Accuracy.  Master's  Thesis,  VPI&SU.  70  p. 
Dunn,  R ,  A.  R.  Harrison,  and  J.C.  White.  1990.  Positional  Accuracy  and 

Measurement  Error  in  Digital  Databases  of  Land  Use:  An  Empirical  Study. 

International  Journal  of  Geographical  Information  Systems.  4(4):385-398. 
Evans,  D.L.,  R.W  Carraway,  and  G.T.  Simmons.  1992.  Use  of  Global 

Positioning  System  (GPS)  for  Forest-Plot  Location.  Southern  Journal  of 

Applied  Forestry.  16(2):67-70. 
Fernandez,  N.,  D.  F.  Lozano-Garcia,  G.  Deeds,  and  C.  J.  Johannsen.  1991 

Accuracy  Assessment  of  Map  Coordinate  Retrieval.  Photogrammetric 

Engineering  &  Remote  Sensing  57(1 1):  1447-1452. 
Ran,  L.,  1992.  Single  Digital-Photo  Correction  for  a  GIS  Application  and  Error 

Analysis.  Master's  Thesis,  VPI&SU.  139  pp. 
Walsh,  S.  J.,  D.  Lightfoot,  D.  Butler.  1987.  Recognition  and  Assessment  of 

Error  in  Geographic  Information  Systems.  Photogrammetric  Engineering  and 

Remote  Sensing.  53(10):  1423-1430. 
Warner,  W.  S.,  and  W.  Carson.  1991.  Errors  Associated  with  a  Standard 

Digitizing  Tablet.  ITC  Journal.  No.  2:82-85  pp. 
Wiles,  S.  J.  1988.  Evaluation  of  Photographic  Properties  for  Area  Estimation. 

Master's  Thesis.  VPI&SU.  100  p. 


BIOGRAPHICAL  SKETCH 


Russell  G.  Combs,  Jr.  is  a  Database/GPS  Forester  with  Canal  Forest  Resources, 
Inc.  in  Charlotte,  NC.  Russ  holds  a  B.S.  in  forestry  and  a  M.S.  in  GIS  and  GPS 
from  Virginia  Tech.  Russ  has  been  employed  with  CFR  for  the  past  year  where  he 
is  responsible  for  managing  the  inventory  database  for  CFR's  clients  as  well  as 
heading  up  Canal's  company  wide  GPS  program  and  assisting  on  various  GIS 
projects  and  analysis. 

James  L.  Smith  holds  a  B.S.  in  Forestry  and  an  M.S.  in  Forest  Biometrics  from 
the  University  of  Georgia,  and  a  Ph.  D.  in  Forest  Biometrics  and  Remote  Sensing 
from  Virginia  Tech.  He  was  a  member  of  the  faculty  of  the  Virginia  Tech 
Department  of  Forestry  at  Virginia  Tech  for  13  years,  where  he  specialized  in 
quantitative  GIS  and  remote  sensing  issues.  For  the  last  two  years,  Jim  has  been 
the  Manager  of  the  GIS  and  Remote  Sensing  Group  at  Canal  Forest  Resources, 
Inc.  in  Charlotte,  NC. 

Paul  V.  Bolstad  is  an  Assistant  Professor  in  the  Department  of  Forest  Resources 
at  the  University  of  Minnesota.  Paul  holds  a  B.S.  in  Forestry  from  the  University 
of  California  at  Berkeley,  a  M.S.  in  Forestry  from  NC.  State,  and  a  Ph.  D.  in 
Forestry  from  Wisconsin  University.  Paul  was  a  member  of  the  faculty  of  the 
Virginia  Tech  Department  of  Forestry  at  Virginia  Tech  for  6  years,  where  he 
specialized  in  GIS,  Remote  Sensing,  and  GPS  issues. 


365 


Large  Scale  Tropical  Forest  Change 
Monitoring  Using  Multiple  Resolution 
Satellite  Data  :  from  Hot  Spot  Detection 

to  Global  Deforestation  Assessment  ? 

Herve  Jeanjean1,  Frederic  Achard2  and  Jean-Paul  Malingreau3 

Abstract  -  The  rate  of  tropical  forest  degradation  and  deforestation  is  a  serious 
threat  on  environmental  issues,  socio-economic  balances  and  land  use  planning 
and  is  considered  as  a  major  concern  for  decision-makers  at  global,  regional 
and  local  scales.  Appropriate  tools  are  needed  to  address  this  multiscale  and 
multipurpose  issue.  Sampling  techniques  can  be  applied  in  large  forest  surveys 
using  high  resolution  satellite  data.  "Wall  to  wall"  coverage  is  a  feasible 
alternative,  but  due  to  the  high  cost  of  data,  low  resolution  satellite  data  are 
more  adapted  to  this  method,  provided  that  estimations  are  corrected  using 
multistage  procedures.  Recent  advances  have  proved  that  this  approach  is 
feasible  for  mapping  and  assessing  tropical  forest  on  large  scale  with  reasonable 
accuracy.  However,  for  forest  change  detection  and  monitoring,  few  methods 
have  been  developed  so  far  enabling  the  estimate  of  change  at  global  scale.  In 
this  paper,  change  detection  based  on  a  combination  of  LRSD  such  as  AVHRR 
data  and  HRSD  such  as  TM  and  SPOT  data  is  addressed.  A  preliminary 
investigation  is  carried  out  in  a  pilot  area  in  Southeast  Asia  where  changes 
detected  with  HRSD  (hot  spot  areas)  are  compared  with  AVHRR  forest  class 
proportions  and  fragmentation  patterns.  A  discussion  is  then  proposed  on  the 
necessary  steps  towards  an  operational  system  for  global  tropical  forest 
monitoring. 


INTRODUCTION 

Tropical  forests  constitute  fragile  ecosystems  which  shelter  most  of  the  world 
animal  and  vegetal  species.  Playing  an  essential  role  in  global  environmental  and 
climatic  balance,  tropical  forests  are  threatenend  by  a  fast  disappearance  whith  dramatic 
consequences  on  biodiversity  loss,  shortage  of  wood  and  non-wood  products,  soil 
degradation  and  erosion,  deterioration  of  hydrological  systems,  disruption  of  local  socio- 
economic development... 

According  to  the  last  estimates  of  FAO  FRA  1990,  15.4  millions  ha  of  tropical 
forests  have  disappeared  each  year  between  1980  and  1990,  which  represent  an  average 
deforestation  rate  of  0.8  %  of  total  tropical  forest  resources.  In  1980,  tropical  forests 
were  estimated  at  1,910  millions  ha,  and  1,756  millions  ha  in  1990.  The  deforestation 
mechanism  has  become  a  critical  issue  not  only  for  the  countries  where  this  problem  is  a 
real  concern,  but  also  for  the  international  scientific  community  who  has  launched 


1  Remote  sensing  and  GIS  expert  in  forestry,  SCOT  Conseil,  1  rue  Hermes,  F-31526  Ramonville  France 

2  MTV  Group,  TREES  Project  Manager,  Institute  of  Remote  Sensing  Applications,  1-21020  Ispra  Italy 

3  Head  of  Monitoring  Tropical  Vegetation  Unit,  TREES  Project  Coordinator,  Institute  of  Remote  Sensing  Applications, 
1-21020  Ispra  Italy 


366 


important  research  programmes  to  better  evaluate  its  exact  magnitude  and  to  improve  the 
knowledge  of  the  different  factors  interacting  in  this  complex  processus.  Intensive 
logging  practices,  shifting  cultivations  and  conversion  of  forest  lands  to  agricultural 
areas  are  some  of  the  main  causes,  direct  or  undirect,  of  deforestation.  Above  all,  the 
modes  of  appropriation  of  forest  lands  by  local  farmers  represent  in  many  cases  a  critical 
issue  which  interacts  with  legal  aspects  of  land  management.  To  tackle  the 
environmental  effects  of  deforestation,  one  of  the  necessary  actions  consists  in  building 
up  a  continuous  monitoring  systems  along  three  major  steps  :  (1)  development  of  sound 
methods  for  estimating  forest  areas  (static  approach),  (2)  development  of  sound  methods 
for  estimating  forest  cover  changes  on  a  quantitative  and  structural  basis  (dynamic 
approach)  and  (3)  modeling  of  deforestation  processes  (anticipating  approach). 
Considerable  advances  have  been  made  on  the  development  of  methods  for  retreiving 
forest  proportions  based  on  CRSD.  Few  methods  have  been  developed  for  change 
detection  and  quantification,  and  they  are  generally  using  sampling  techniques  with 
HRSD.  The  modeling  of  deforestation  requires  to  consider  many  parameters  which 
encompass  not  only  the  vegetation  characteristics  of  the  area,  but  also  physical  and 
anthropic  characteristics. 

This  paper  is  mainly  addressing  the  second  issue,  after  reviewing  some  techniques 
for  global  estimates  of  forest  areas. 

GLOBAL  ESTIMATE  OF  FOREST  COVER  PROPORTION 

The  main  actors 

As  far  as  global  assessment  is  concerned,  CRSD  such  as  AVHRR  data  can  provide 
valuable  information  for  forest  resources  assessment.  Future  sensors  (vegetation 
instrument  on  board  SPOT  4  and  MODIS)  are  expected  to  give  further  possibilities  in 
this  field. 

Several  institutions  or  programmes  have  already  started  significant  activities  in 
global  survey  of  tropical  forests  :  (i)  the  European  Union  with  the  TREES  Project 
(TRopical  Ecosytem  Environment  observation  by  Satellites)  where  wall  to  wall  coverage 
is  made  using  AVHRR  data  with  a  sample  of  HRSD  for  calibration  purposes,  (ii)  FAO 
with  the  Forest  Resources  Assessment  1990  project  where  TM  data  are  sampled  over  the 
entire  tropical  belt,  (iii)  IUFRO  (International  Union  of  Forest  Research  Organizations) 
which  has  published  some  guidelines  for  forest  monitoring  (Paivinen,  1994),  (iv)  NASA 
which  is  carrying  out  several  projects  in  South  America  and  Africa  (Landsat  Pathfinder 
project  in  the  framework  of  the  Global  Change  Research  Programme  :  Chomentowski  et 
al.,  1994),  (v)  UNEP/GRID  which  has  undertaken  some  global  analysis  with  AVHRR 
data  in  the  amazon  basin,  and  has  been  involved  with  TREES  in  Southeast  Asia,  (vi)  the 
Woods  Hole  Research  Centre  which  has  realized  a  map  of  South  America  using 
different  sensors,  (vii)  IUCN  (International  Union  for  Conservation  of  Nature  and 
Natural  Resources)  and  WCMW  (World  Conservation  and  Monitoring  Centre)  which 
published  in  the  1990's  the  Conservation  Atlas  of  Tropical  Forests,  (viii)  IGBP 
(International  Global  Biosphere  Programme)  which  is  working  on  global  change  issues. 
Over  Africa,  other  regional  initiatives  with  environmental  components  are  in  progress  or 
in  preparation  :  the  CARPE  (Central  African  Regional  Program  for  Environment),  and 
RIMP  (Regional  Information  Management  Programme). 


367 


Existing  methods  for  global  forest  survey 


Four  different  methods  can  be  applied  for  estimating  tropical  forest  areas  on  a 
global  basis  (Czaplewski,  1992).  A  first  approach  is  consisting  in  compiling  forest 
statistics  provided  by  national  or  regional  forest  inventories.  FAO  used  this  approach  in 
1980  and  to  a  limited  extent  in  1990.  The  main  problems  of  this  method  are  the 
discrepancies  between  the  nomenclatures,  the  variables,  the  accuracy  in  the  sampling 
designs,  the  dates  of  the  inventories.  A  second  method  is  based  on  the  analysis  of  a  wall 
to  wall  coverage  of  CRSD  data  (AVHRR  data).  In  this  approach,  forest  resources  can  be 
mapped,  e.g.  located,  and  estimated  using  some  AVHRR  classification  schemes. 
However,  the  estimation  of  forest  proportion  is  presenting  a  systematic  bias  due  to 
aggregation  effects.  The  magnitude  of  this  error  depends  on  three  factors  :  the  spatial 
resolution  of  the  map,  the  initial  proportions  of  forest  in  the  landscape,  and  the  spatial 
arrangement  of  forest  at  fine  resolution  (Turner  et  al.,  1989;  Moody  and  Woodcock, 
1994).  It  is  therefore  recommended  to  apply  a  correction  function  which  may  be  derived 
from  a  double  sampling  design  with  regression  estimator.  The  regression  is  performed 
between  the  auxiliary  variable,  measured  from  the  AVHRR  data  set,  and  the  target 
variable,  measured  on  a  sample  of  HRSD  sites.  This  method  has  been  successfully 
implemented  by  the  TREES  project  (Mayaux  et  Lambin,  1995).  The  third  method  is 
aimed  at  estimating  forest  proportion  from  a  sample  of  high  resolution  images  :  this 
method  has  been  tested  and  implemented  by  FAO  in  the  Forest  Resources  Assessment 
1990.  Moreover,  if  several  images  are  acquired  on  each  site,  it  is  possible  to  estimate  the 
deforestation  rate.  FAO  has  estimated  the  deforestation  rate  between  1980  and  1990.  In 
the  fourth  approach,  national  mapping  results  are  compiled  at  regional  or  global  level  : 
as  for  the  first  method,  this  approach  is  limited  by  the  discrepancies  in  nomenclature, 
scales  and  dates. 

The  multi  scale  approach 

The  second  method,  based  on  CRSD  such  as  AVHRR  data,  is  the  only  one 
providing  some  estimates  of  global  forest  resources  and  a  location  of  those  resources  at 
the  same  time.  To  increase  the  accuracy  of  the  estimate,  many  investigators  suggested  to 
use  HRSD  as  reference  data  (Tucker  and  al.,  1984;  Nelson  and  Holben,  1988;  Woodwell 
and  al.,  1987;  Malingreau  and  Tucker,  1988;  Paivinen  and  Witt,  1988;  Malingreau  and 
al.,  1989;  Stone  and  al.,  1990;  Nelson  and  Horning,  1993).  The  main  advantage  of  this 
approach  is  the  daily  availability  of  data,  and  the  spatial  resolution  (1.1  Km)  which  is 
more  adapted  to  the  scale  of  study.  However,  using  coarse  resolution  satellite  data  leads 
to  a  loss  of  details  which  depends  on  the  spatial  structure  of  the  landscape  (Woodcock 
and  Strahler,  1987;  Townshend  and  Justice,  1988).  The  integration  of  high  resolution 
satellite  data  represents  a  double  advantage  with  the  possibility  to  first  validate  the  broad 
scale  maps  by  comparing  fine  scale  observations  with  AVHRR  classification  and  second 
to  develop  a  correction  function  with  respect  to  proportional  errors  in  order  to  estimate 
tropical  forest  areas  from  coarse  resolution  classifications. 

Building  a  correction  function  :  the  lessons  of  the  TREES  project 

The  Joint  Research  Centre  TREES  project  is  aimed  at  mapping  and  estimating 
tropical  forest  resources,  detecting  and  monitoring  changes  and  modeling  deforestation 


368 


(Malingreau  and  al.,  1993).  A  correction  model  for  proportional  errors  has  been  tested 
and  developed.  To  calibrate  broad  scale  tropical  forest  area  estimates,  an  inverse 
calibration  model  was  applied  (Brown,  1982;  Czaplewski,  1992)  by  integrating  in  the 
double  sampling  approach  a  measure  of  the  spatial  fragmentation  of  the  forest-non  forest 
interface  (Mayaux  and  Lambin,  1995).  The  inversion  model  can  be  expressed  as  (Eq.  1): 

Pf,  =f[PCi,SCi]  (1) 

where  PFj  is  the  forest  proportion  at  fine  scale,  Pcj  the  forest  proportion  at  broad  scale, 

and  SCj  a  descriptor  of  the  spatial  pattern  of  the  class  i  at  coarse  resolution.  Several 

tests  have  shown  that  the  Matheron  index  (Kleinn  et  al.,  1993)  is  a  good  descriptor  of 
the  spatial  structure  in  AVHRR  classifications.  It  is  defines  as  (Eq.  2)  : 

number  of  runs  between  forest  and  other  cover  type  pixels 

M  =  .  ,  —  (2) 

^/(number  of  forest  pixels) y(total  number  of  pixels) 

The  two-step  correction  model  for  aggregation  errors,  in  which  forest  cover 
proportion  at  fine  and  coarse  resolution  as  well  as  the  fragmentation  index  are  calculated 
in  13x13  AVHRR  pixels  blocks,  was  developed  with  two  major  regression  phases  :  first, 
a  linear  regression  was  performed  within  each  equal-size  subset  of  similar  spatial  pattern 
between  forest  cover  proportions  at  fine  and  coarse  resolution,  and  second,  a  regression 
was  performed  between  the  fragmentation  measure  and  the  intercept  and  the  slope  of  the 
previous  regressions.  This  calibration  procedure  improved  by  25.5  %  the  reliability  of 
the  retrieval  of  forest  cover  proportions  from  coarse  resolution  data  by  comparison  to  a 
simple  correction  function  relating  directly  proportions  at  coarse  and  fine  resolutions 
(Mayaux  and  Lambin,  1995).  Further  improvements  are  in  progress,  with  the 
introduction  of  spectral  mixture  models  at  coarse  resolution  scale.  The  integration  of 
spatial  information  in  the  mixed  pixel  estimator  improved  significantly  the  reliability  of 
the  model  (Mayaux  and  Lambin,  1996). 

FOREST  CHANGE  DETECTION 

Digital  change  detection  methods 

Forest  ecosystems  are  in  continuous  evolution.  Changes  can  be  defined  as  "an 
alteration  in  the  surface  components  of  the  vegetation  cover"  (Milne,  1988)  or  as  "a 
spectral/spatial  movement  of  a  vegetation  entity  over  time"  (Lund,  1983).  The  intensity 
of  change  can  be  abrupt  (clear  cutting,  deforestation,  and  forest  fires)  or  diffuse  and 
gradual  (stands  growth).  Various  classifications  of  forest  change  have  been  proposed  by 
Aldrich  (1975)  who  suggested  nine  general  forest  disturbance  classes,  by  Col  well  (1980) 
who  proposed  a  more  hierarchical  framework,  and  by  Hame  (1986)  who  took  a  more 
mechanistic  view  of  the  problem.  Ecological  aspects  have  been  then  developed  by 
Hobbs  (1990),  and  Khorram  (1994)  who  concentrated  more  on  the  spatial  environment 
in  which  the  change  occurs. 

Critical  issues  in  forest  change  detection  are  the  appropriate  selection  of  images,  the 
choice  of  sensors,  the  change  categories  and  the  change  detection  algorithms  (Coppin, 
1994).  Moreover,  the  change  interval  length,  e.g.  the  temporal  resolution,  is  a  key  aspect 
in  change  detection.  The  accuracy  of  forest  cover  change  information  is  closely  related 
to  the  maximization  of  the  signal-to  noise  ratio.  When  this  ratio  is  not  optimum,  unreal 


369 


changes  may  be  detected,  and  are  caused  by  differences  in  scattering  conditions, 
differences  in  water  content  in  the  atmosphere,  variations  in  the  solar  azimuth  angles, 
and  sensors  calibration  inconsistencies  (Hade,  1988).  Most  digital  change  detection 
methods  are  based  on  per-pixel  classifiers  (Teuber,  1990).  But  the  decision  rules,  which 
comprise  the  complete  range  of  pattern  recognition  techniques,  constitute  the  most 
sensitive  aspect  in  any  change  detection  model.  Numerous  detection  algorithms  have 
been  developed  during  the  past  years.  They  can  be  grouped  into  two  main  categories  : 
change  measurement  methods,  and  classification  schemes  (Malila,  1980;  Pilon  and  al., 
1987).  It  is  also  possible  to  separate  the  simultaneous  analysis  of  multitemporal  data 
versus  comparative  analysis  of  independent  single  date  classifications  (Singh,  1989). 
Change  detection  methods  may  be  characterized  by  two  main  phases  :  first  the  data 
transformation  procedure,  and  second  the  analysis  techniques  applied  to  delineate  the 
changed  areas  (Nelson,  1983).  The  transformation  procedure  can  be  no  transformation 
(raw  data),  images  differencing,  bands  ratioing,  vegetation  index  differencing,  regression 
between  bands,  multitemporal  linear  data  transformation  (principal  component  analysis 
and  tasseled  cap),  change  vector  analysis,  and  comparison  of  single  date  classifications. 
The  analysis  techniques  used  to  detect  change  are  usually  deriving  from  thresholding, 
supervised  or  unsupervised  classification,  spatial  analysis,  or  layered  spectral/temporal 
analysis.  For  a  review  of  change  detection  techniques,  see  Singh  (1989)  and  Coppin 
(1994). 

The  multi  scale  approach  in  change  detection 

How  to  conciliate  broad  scale  and  fine  scale  observations 

Most  of  the  detection  change  studies  have  used  high  resolution  satellite  data  since 
the  changes  are  easier  to  detect  at  fine  scale.  At  broad  scale,  change  vector  analysis 
techniques  have  been  applied  to  characterize  vegetation  changes  in  Africa  (Lambin  and 
Strahler,  1994).  Very  few  studies  have  been  carried  out  using  several  resolution  satellite 
data.  As  far  as  vegetation  monitoring  at  global  scale  is  concerned,  there  is  a  need  to  get 
observations  at  different  scales  both  in  temporal  and  spatial  terms  (Malingreau  and 
Bel  ward,  1992).  A  relationship  between  forest  clearing  rate  based  on  MSS-TM  analysis 
and  fire  activity  based  on  AVHRR  interpretation  was  studied  in  Mato  Grosso,  Brazil 
(Nelson  and  al.,  1987).  The  results  suggested  to  use  AVHRR  data  as  a  stratification  tool 
rather  than  as  a  auxiliary  variable  in  the  sampling  design.  Townshend  and  Justice  (1995) 
have  looked  into  the  spatial  frequencies  of  two  MSS  single  date  images  and  of  the 
change  image  obtained  by  substracting  the  NDVI.  Considerable  variability  in  the  spatial 
frequency  was  found  on  the  two  images,  with  a  spatial  resolution  ranging  from  125  m  to 
64  Km.  But  no  relationship  was  detected  between  the  spatial  frequencies  of  the  two 
single  date  images  and  the  change  image. 

To  estimate  changes,  a  simple  approach  can  be  applied  consisting  in  substracting 
one  map  (or  classification)  from  another.  But  the  bias  and  the  resulting  variance  are 
unknown  and  might  be  substantial.  Statistical  procedures  were  tested  for  estimating 
change  from  thematic  maps  available  for  the  same  region  at  two  times  (Van  Deusen, 
1994).  Subsampling  was  used  to  obtain  estimates  of  the  true  class  of  some  pixels  at  each 
of  the  two  times.  Simulation  results  verified  that  the  procedures  produce  unbiased 
estimates  and  that  associated  variance  estimates  are  reliable.  Van  Deusen  pointed  out 
that  this  statistical  approach  could  be  applied  on  low  resolution  thematic  maps  by  using 


370 


subsamples  of  high  resolution  data  to  refine  the  true  class  proportions.  In  a  global  forest 
monitoring  exercise,  Beltz  and  al.  (1992)  recommended  to  use  multi-stage  sampling  with 
sampling  units  proportional  to  forest  cover  or  deforestation,  and  sampling  with  partial 
replacement. 

A  correction  function  for  change  detection 

As  for  the  estimate  of  forest  areas,  an  inverse  correction  model  can  be  built  to  refine 
the  estimate  of  forest  cover  changes.  A  simple  model  can  be  built  by  relating  the 
estimate  of  change  at  broad  scale  with  the  estimate  of  change  at  fine  scale  (Eq.  3  )  : 


where  APFj  and  APCj  are  the  estimates  of  changes  at  fine  scale  and  coarse  scale  of 

class  i  between  time  ti  and  t2.  The  introduction  of  the  spatial  structure  in  the  model 
would  lead  to  the  following  equation  (Eq.  4)  : 


of  the  time  inteval  during  which  the  change  in  forest  proportion  will  be  estimated  :  this  a 
priori  knowledge  of  the  spatial  arrangement  of  forest  cover  can  be  considered  as  an 
indicator  of  future  deforestation.  A  fragmented  pattern  with  many  forest-non  forest 
contacts  may  be  more  sensitive  to  deforestation  than  a  continuous  block  of  forest  cover. 
However,  this  indicator  of  future  deforestation  and  the  magnitude  of  change  are  very 
much  depending  on  the  type  of  deforestation,  e.g.  logging  activities,  clear  cutting, 
shifting  cultivations. ..etc.  On  the  contratry,  SCj(t2) characterizes  the  spatial  structure  at 

the  end  of  the  interval,  and  thus  can  be  considered  as  an  indicator  of  past  deforestation 
activities.  When  CRSD  data  are  not  available  at  time  ti  and  t2,  the  correction  model  can 
be  simplified  as  follows  (Eq.  5),  whith  tj  being  between  ti  and  t2 : 


Preliminary  results  on  a  test  site 

A  study  was  carried  out  on  a  test  area  located  in  Vietnam  where  deforestation  is 
very  active,  and  field  observations  were  available.  An  AVHRR  classification  from  1992- 
93  and  a  TM  scene  (124/52  from  December  30,  1990)  were  provided  by  the  TREES 
project.  A  SPOT  image  was  purchased  (KJ  277-326  from  February  25,  1995).  The 
acquisition  periods  of  the  TM  and  SPOT  scenes  correspond  to  the  dry  season.  Geometric 
corrections  were  performed  between  all  satellite  data  with  a  common  projection  system. 
The  total  RMS  between  the  SPOT  scene  and  the  TM  scene  was  0.13  pixel.  The  RMS 
obtained  between  AVHRR  data  and  the  TM  scene  was  0.11  AVHRR  pixel.  A 
multitemporal  unsupervised  classification  was  performed  using  the  three  bands  of  SPOT 
and  the  first  five  bands  of  TM.  The  classes  were  then  grouped  into  four  classes  : 
unchanged  forest,  unchanged  non  forest,  deforested  areas  (between  1990  and  1995),  and 
clouds  or  shadows.  The  AVHRR  classification  is  giving  three  main  classes  :  dense  forest 


APf,  =f[APc4] 


(3) 


(5) 


371 


(forest  cover  percentage  over  70  %),  fragmented  forest  (forest  cover  percentage  between 
40  and  70  %),  and  non  forest  (forest  cover  percentage  below  40  %).  Forest  cover 
proportions  were  calculated  in  9x9  blocks  of  AVHRR  pixels. 

Figure  1 .  Relationship  between  the  deforestation  rate  at  fine  scale  and  the 
proportion  of  AVHRR  fragmented  forest. 


30  T 


The  relationship  between  the  deforestation  rate  measured  at  fine  scale  and  the  proportion 
of  AVHRR  fragmented  forest  (figure  1)  shows  a  rather  good  correlation  (coefficient  of 
determination  is  0.61).  It  can  be  noticed  that  the  deforestation  rate  increases  with  the 
proportion  of  fragmented  forest,  but  this  trend  stops  at  a  proportion  of  50  %.  A  similar 
trend  is  observed  with  the  Matheron  index,  with  a  less  good  correlation  with  the 
deforestation  rate  (r  =  0.33).  On  the  contrary,  when  AVHRR  dense  forest  increases,  the 
deforestation  rate  decreases.  It  can  be  concluded  that  the  deforestation  rate  is  in  this  case 
governed  by  the  fragmentation  pattern  observed  at  broad  scale.  The  AVHRR 
classification  and  the  TM/SPOT  classification  are  presented  in  figure  2. 

Figure  2.  Left  :  multitemporal  classification  TM(1990)-SPOT(1995)  with  unchanged  forest 
(black),  unchanged  non  forest  (white)  and  deforestated  areas  (grey).  Right  :  AVHRR 
classification  (1993)  with  dense  forest  (black),  fragmented  forest  (grey)  and  non  forest 
(white). 


372 


It  can  be  concluded  that  there  is  a  consistent  relationship  between  the  estimate  of 
deforestation  at  fine  scale  and  forest  cover  parameters  at  broad  scale.  The  AVHRR 
fragmented  forest  class  is  already  containing  some  information  of  the  degree  of 
fragmentation  within  the  pixel.  The  fragmentation  pattern  measured  with  the  Matheron 
index  in  a  block  of  pixels  is  likely  to  improve  the  correction  model.  However,  more 
HRSD  sites  are  needed  to  test  the  model  and  its  real  performance.  Moreover,  it  is 
suggested  to  develop  a  more  complete  model  integrating  the  estimate  of  forest  cover 
change  at  broad  scale  (eq.  4).  For  this  purpose,  a  more  complete  set  of  AVHRR  data  is 
required.  Phase  2  of  the  TREES  project  which  has  just  been  launched  should  focus  on 
global  forest  cover  change  with  the  acquisition  and  processing  of  new  AVHRR  data  set. 
It  can  be  also  pointed  out  that  the  correction  model  should  pave  the  way  for  a  global 
deforestation  model  with  the  integration  of  auxiliary  parameters  such  as  socio-economic 
data,  population  figures,  infrastructure,  and  local  environment  conditions  (topography, 
soils...).  The  Tropical  Forest  Information  System  (TFIS)  set  up  by  the  TREES  project 
will  be  the  basis  for  reaching  this  final  goal. 

REFERENCES 

Achard  F.,  and  D'Souza  G.,  1994,  Collection  and  processing  of  NOAA  AVHRR  1  km  resolution 

data  for  tropical  forest  resource  assessment,  TREES  Series  A:  Technical  document  n°2,  EUR 

16055  EN,  European  Commission,  Luxembourg,  58p. 
Beltz  R.,  Evans  D.,  Czaplewski  R.L.  and  Van  Deusen  P.,  1992,  National  forest  area  and  rates  of 

change  estimates  using  satellite  data,  Report  of  the  UNEP/FAO  Expert  Consultation  on 

Environmental  Parameters  in  Future  Global  Forest  Assessments. 
Bel  ward  A.S.,  1992,  Spatial  attributes  of  AVHRR  imagery  for  environmental  monitoring,  Int.  J. 

Remote  Sensing,  vol.  13,  N°2,  193-208. 
Chomentowski  W.,  Salas  B.  and  Skole  D.,   1994,  Landsat  Pathfinder  project  advances 

deforestation  mapping,  GIS  World  7(4):  34-38. 
Coppin  P.R.  and  Bauer  M.E.,  July  1994,  Processing  of  multitemporal  landsat  TM  imagery  to 

optimize  extraction  of  forest  cover  change  features,  IEEE  Transactions  on  geoscience  and 

remote  sensing,  Vol.  32,  N°4. 
Coppin  P.R.,  September  1994,  Digital  change  detection  in  forest  ecosystems:  where  are  we  and 

where  are  we  going?,  ISPRS,  Proceedings  "Resource  and  environmental  monitoring",  Brazil. 
Czaplewski  R.L.  and  Catts  G.P.,  1992,  Calibration  of  remotely  sensed  proportion  or  area  estimates 

for  misclassification  error,  Remote  Sens.  Environ.  39:29-43. 
Czaplewski  R.L.,  Analysis  of  alternative  sample  survey  designs,  FAO  1991. 
Food  and  Agriculture  Organization,  1993,  Forest  resources  assessment  1990:  tropical  countries, 

FAO  Forestry  Paper  1 12,  Rome,  61pp. 
Jeanjean  H..  Malingreau  J. P.  and  Achard  F.,  1994,  Tropical  forest  fragmentation:  typology  and 

characterisation,  in  Proceedings  of  the  European  Symposium  on  Satellite  Remote  Sensing, 

26-30  September,  Rome. 
Klein  C,  Dees  M.  and  Pelz  D.R.,  1993,  Sampling  aspects  in  the  TREES  project:  global  inventory 

of  tropical  forests,  Final  report  to  the  Joint  Research  Center,  Contract  n°  5014-92-10  ED  ISP 

D,  Universitat  Freiburg,  36pp. 
Lambin  E.F.  and  Strahler  A.H.,  1994,  Change-vector  analysis  in  multitemporal  space:  a  tool  to 

detect  and  categorize  land-cover  change  processes  using  high  temporal-resolution  satellite 

data,  Remote  Sens.  Environ.  48:231-244. 
Lambin  E.F.  and  Strahler  A.H.,  1994,  Indicators  of  land-cover  change  for  change-vector  analysis 

in  multitemporal  space  at  coarse  spatial  scales,  Int.  J.  Remote  Sensing,  Vol.  15,  N°10,  2099- 

2119. 

Malingreau  J. P.  and  Belward  A.S.,  1992,  Scale  considerations  in  vegetation  monitoring  using 
AVHRR  data,  Int.  J.  Remote  Sensing,  Vol.  13,  N°12,  2289-2307. 


373 


Malingreau  J.P.  and  Tucker  C.J.,  1988,  Large  scale  deforestation  in  the  Southeastern  Amazon 

Basin  of  Brazil,  AMBIO  17:49-55. 
Malingreau  J.P.  and  Tucker  C.J.,  1990,  Cover,  Ranching  in  the  Amazon  Basin,  Large-scale 

changes  observed  by  AVHRR,  Int.  J.  Remote  Sensing,  Vol.  11,  N°2,  187-189. 
Malingreau  J.P.,  Achard  F.,  D'Souza  G.  et  al.,  1995,  AVHRR  for  global  tropical  forest 

monitoring:  the  lessons  of  the  TREES  project,  Remote  Sensing  Reviews  12:  29-40. 
Malingreau  J.P.,  Tucker  C.J.  and  Laporte  N.,  1989,  AVHRR  for  monitoring  global  tropical 

deforestation,  Int.  J.  Remote  Sensing  10:  855-867. 
Mayaux  P.  and  Lambin  E.F.,  1995,  Estimation  of  tropical  forest  area  from  coarse  spatial 

resolution  data:  a  two-step  correction  function  for  proportionnal  errors  due  to  spatial 

aggregation,  Remote  Sensing  Environ.  53:  1-15,  Elsevier  Science  Inc. 
Mayaux  P.  and  Lambin  E.F.,  September  1995,  Improvements  of  the  calibration  of  coarse 

resolution  tropical  forest  area  estimates  with  spatial  texture  measures,  Manuscript  submitted 

to  Remote  Sensing  of  Environment. 
Moody  A.  and  Curtis  E.  Woodcock,  May  1994,  Scale-dependent  errors  in  the  estimation  of  land- 
cover  proportions:  implications  for  global  land-cover  datasets,  Photogrammetric  Engineering 

&  Remote  sensing,  Vol.  60,  N°5,  585-594pp. 
Nelson  R.  and  Holben  Brent,  1986,  Identifying  deforestation  in  Brazil  using  multiresolution 

satellite  data,  Int.  J.  Remote  Sensing,  vol.  7,  N°3,  429-448. 
Nelson  R.,  February  1989,  Regression  and  ratio  estimators  to  integrate  AVHRR  and  MSS  data, 

Remote  Sens.  Environ.  30:201-216. 
Nelson  R.,  Horning  N.  and  Stone  T.A.,  1987,  Determining  the  rate  of  forest  conversion  in  Mato 

Grosso,  Brazil,  using  Landsat  MSS  and  AVHRR  data,  Int.  J.  Remote  Sensing,  Vol.  8,  N°12, 

1767-1784. 

Paivinen   R.,    1994,   IUFRO   International   guidelines   for   forest   monitoring,  Directrices 

internacionales  de  IUFRO  para  la  monitorizacion  de  los  recursos  forestales,  IUFRO  World 

series,  Vol.  5,  ISSN  1016-3263. 
Sader  S.A.  and  Winne  J.C.,  1992,  RGB-NDVI  colour  composites  for  visualizing  forest  change 

dynamics,  Int.  J.  Remote  Sensing,  Vol.  13,  N°16,  3055-3067. 
Singh  A.,  1989,  Digital  change  detection  techniques  using  remotely-sensed  data,  Int.  J.  Remote 

Sensing,  Vol.  10,  N°6,  989-1003. 
Stone  T.A.  and  Schlesinger  P.,  June  1990,  Monitoring  deforestation  in  the  tropics  with  NOAA 

AVHRR  and  Landsat  data. 
Townshend  J.R.G.  and  Justice  CO.,  1988,  Selecting  the  spatial  resolution  of  satellite  sensors 

required  for  global  monitoring  of  land  transformations,  Int.  J.  Remote  Sensing,  Vol.  9,  N°2, 

187-236. 

Townshend  J.R.G.  and  Justice  CO.,  1995,  Spatial  variability  of  images  and  the  monitoring  of 
changes  in  the  Normalized  Difference  Vegetation  Index,  Int.  J.  Remote  Sensing,  Vol.  16, 
N°12,  2187-2195. 

Van  Deusen  P.C,  1994,  Correcting  bias  in  change  thematic  maps,  Remote  Sens.  Environ.  50:67- 
73. 

Woodcock  CE.  and  Strahler  A.H.,  1987,  The  factor  of  scale  in  remote  sensing,  Remote  Sens. 
Environ.  21:311-332. 

Zhu  Z.,  1994,  Forest  density  mapping  in  the  lower  48  States:  a  regression  procedure, 
U.S.  Department  of  Agriculture,  Forest  Service,  Southern  Forest  Experiment  Station,  New 
Orleans,  LA,  Research  Paper  SO-280,  9pp. 


374 


Optimum  Area  Sampling  Frame  Using 
High  Resolution  Satellite  Images  with 
Operational  Objective  : 
How  to  Conciliate  Statistical 
Requirements  and  Practical  Aspects? 

Helene  de  Boissezon1,  Herve  Jeanjean2 


Abstract:  If  high  resolution  satellite  images  can  be  used  for  measuring 
land  use  characteristics,  it  is  most  often  impossible  to  envisage  an 
exhaustive  coverage  of  large  areas.  Sample  survey  methods  represent  a 
feasible  alternative  when  assessing  land  use  surface.  In  many  cases, 
classical  sampling  survey  techniques  can  be  used.  However,  when  spatial 
correlation  is  observed  between  sites,  geostatistics  principles  could  be 
introduced  into  the  inventory  design.  Using  classical  statistical  techniques, 
the  target  variable  has  to  be  estimated  from  the  elementary  measurements 
on  the  sample  sites  through  an  extrapolation  model.  When  designing  such 
statistical  process,  many  issues  have  to  be  addressed,  such  as  the 
characteristics  of  the  population  to  be  considered,  the  sampling  design 
itself,  the  extrapolation  model  to  be  applied  on  the  whole  study  area,  and 
the  sampling  error.  Practical  considerations  and  constraints  should  not  be 
underestimated  in  any  sample  survey.  In  this  paper,  statistical  issues  on 
sampling  frame  design  based  on  high  resolution  satellite  images  are 
'addressed  with  emphasis  on  practical  implementation.  Sampling  survey 
designs  of  some  operational  remote  sensing  projects  in  the  world  are 
reviewed,  such  as  FAO/FRA  90,  and  the  European  MARS,  TREES, 
FIRS  programmes.  Some  recommandations  are  then  proposed  for  setting 
up  sound  sampling  schemes. 


INTRODUCTION 

In  Earth  observation,  the  analysis  and  monitoring  of  land  surface 
characteristics  such  as  proportion  of  crops  or  forest  types  are  often  requiring  high 
resolution  satellite  images.  Although  some  programmes  are  using  this  type  of  data 
on  a  wall  to  wall  basis,  it  becomes  difficult  to  assess  large  areas  with  a  full 


Remote  sensing  expert  in  agronomic  statistics,  SCOT  Conseil,  1  rue  Hermes,  F-3 1 526  Ramonville  France 

2 

Remote  sensing  expert  in  forestry,  SCOT  Conseil,  1  rue  Hermes,  F-3 1 526  Ramonville  France 


375 


coverage,  because  of  the  high  cost  induced  and  the  huge  amount  of  data  to  be 
processed.  A  sampling  scheme,  based  on  sound  statistical  concepts,  represents  a 
feasible  method.  Some  operational  regional  and  international  programmes  have 
already  developed  and  implemented  such  procedures.  However,  the  sampling 
scheme  has  been  sometimes  designed  a  posteriori,  e.g.  the  samples  have  not  been 
selected  and  located  under  rigourous  rules  at  the  beginning.  Moreover,  pratical 
constraints  can  be  either  overestimated  or  underestimated,  whether  they  are  given 
too  much  or  too  little  consideration.  This  situation  results  in  problems  in  final 
accuracy  and  in  implementing  an  extrapolation  model.  This  paper  is  addressing  the 
optimization  of  an  optimum  area  sampling  frame  design  :  consideration  will  be 
given  to  both  statistical  requirements  and  practical  constraints. 

RECENT  ADVANCES  IN  SAMPLING  DESIGNS 

Few  investigations  have  been  carried  out  on  the  most  adapted  statistical 
procedures  for  covering  large  areas  with  a  sample  of  high  resolution  images. 
However,  the  litterature  is  rather  rich  in  research  studies  concerning  small  areas 
within  satellite  images,  where  sampling  schemes  can  be  tested  and  evaluated  (the 
true  variance  can  be  calculated,  and  the  sampling  can  be  repeated  as  many  times  as 
wishes). 

The  geostatistical  approach 

Geostatistical  concepts  constitute  an  alternative  to  classical  sampling 
techniques.  They  assume  that  the  sample  units  are  not  independent,  and  that  the 
autocorrelation  can  be  used  in  the  model.  Geostatistics  requires  a  systematic 
sampling  rather  than  a  random  sampling.  The  semivariogram  is  the  basic  tool  in 
geostatistics.  In  remote  sensing  images,  strong  correlation  has  been  observed 
between  landscapes  (Ramstein  and  Raffy,  1989;  Webster  and  al.,  1989;  Gohin  and 
Langlois,  1993;  Wood  and  al.,  1994).  McGwire  and  al.  (1993)  has  shown  that  the 
spatial  correlation  is  increasing  with  the  sample  unit  size.  Atkinson  (1995)  confirms 
this  result  and  has  pointed  out  that  the  gain  of  a  systematic  sampling  by 
comparison  to  a  random  sampling  is  increasing  with  the  size  of  the  sample  unit.  A 
minimum  distance  between  sample  sites  can  reduce  the  effects  of  the  spatial 
correlation  (the  sampling  units  become  less  dependent),  and  thus  can  improve  the 
efficiency  of  the  scheme.  According  to  McGwire  and  al.  (1993),  geostatistics  may 
be  tested  and  applied  in  wide  vegetation  mapping  and  monitoring  projects.  The 
studies  carried  out  by  Jupp  and  al.  (1988)  constitute  a  good  starting  point  for 
developing  such  methods.  In  spite  of  this  suggestions,  no  complete  study  has 
proved  the  applicability  of  geostatistical  concepts  when  sampling  large  areas 
(country  to  continent  scale)  with  high  resolution  images.  In  particular,  the  spatial 
correlation  between  SPOT  or  TM  scenes  selected  on  a  systematic  basis  has  still  to 
be  proved.  The  main  bottleneck  for  evaluating  the  usefulness  of  geostatistical  bases 


376 


is  related  to  the  important  means  (number  of  satellite  scenes  and  processing  time) 
which  would  be  required  for  building  and  testing  a  semivariogram.  Therefore, 
classical  statistical  concepts  are  generally  still  prevailing  since  their  application 
rules  and  their  limits  are  well  known. 

Gain  of  systematic  sampling 

Previous  studies  have  shown  that  systematic  sampling  is  becoming  more 
efficient  when  the  sample  size  or  the  sampling  intensity  increases  (Atkinson,  1995). 
In  fact,  as  the  sample  size  increases,  the  sampling  units  are  likely  to  be  more  and 
more  correlated  since  they  are  closer  to  each  other.  However,  it  is  always 
recommended  to  carefully  examine  the  spatial  behaviour  of  the  variable  to 
estimate.  If  the  variable  is  following  a  periodicity  (sequence  of  ridges  and  valleys 
for  instance),  a  bias  is  likely  to  be  introduced.  If  several  simulations  confirmed  the 
gain  of  systematic  sampling  (Cochran,  1981;  Brion  and  Fournier,  1995),  the 
magnitude  of  the  gain  is  varying  and  is  dependending  on  the  spatial  correlation 
(Dunn  and  Harrison,  1993). 

Multi  phase/stage  sampling 

When  the  population  to  be  assessed  in  not  well  known,  e.g.  when  stratification 
is  not  possible,  it  may  be  convenient  to  proceed  to  a  multiple  phase  sampling.  In 
remote  sensing,  a  double  phase  sampling  with  regression  has  been  tested  or 
recommended  in  many  studies,  whether  between  high  resolution  images  and 
ground  measurements  (Smiatek,  1995)  or  between  coarse  resolution  satellite 
images  and  high  resolution  satellite  images  (Mayaux  and  Lambin,  1995;  Kleinn  and 
al.,  1993;  Paivinen  and  Pitkanen,  1992;  Iverson  and  al.,  1989;  Nelson  and  Holben, 
1986).  Multistage  sampling  is  a  feasible  alternative.  In  that  case,  the  total  variance 
is  often  mainly  due  to  the  first  stage. 

Gain  of  stratification 

The  a  priori  stratification  consists  in  partitioning  the  population  with  criteria 
well  correlated  to  the  variable  to  estimate.  The  final  gain  is  depending  on  this 
relationship.  The  a  posteriori  stratification,  which  consists  in  grouping  sampling 
units  in  ensembles  defined  after  the  sampling  scheme,  may  provide  good  results 
(Cochran,  1981;  Dunn  and  Harrison,  1983). 

EXPERIENCE  OF  SOME  LARGE  SCALE  REMOTE  SENSING  PROJECT 

Several  operational  remote  sensing  projects  have  set  up  sound  sampling 
schemes  based  on  high  resolution  satellite  images  as  sampling  units.  The  activity  B 
of  the  MARS  project  (Monitoring  Agriculture  by  Remote  Sensing)  is  an  European 


377 


programme  carried  out  by  the  Joint  Research  Centre  at  the  Institute  of  Remote 
Sensing  Applications  and  is  aimed  at  providing  fast  estimates  of  crops  areas  and 
yields  to  DGVI  in  Brussels.  A  total  of  60  sampling  sites  (40x40  Km2)  have  been 
located  after  stratification,  which  corresponds  to  6  %  sampling  intensity.  The 
allocation  of  sites  is  proportional  to  the  agricultural  lands  in  the  country,  and  is 
coinciding  with  SPOT  scenes  (KJ).  A  weighting  procedure  is  then  employed  to 
derive  the  extrapolation  model.  The  MERA  95  project  (MARS  extension  and 
Environment  Related  Applications)  is  an  extension  of  MARS  to  Central  and 
Eastern  Europe.  The  sampling  design  has  been  slightly  modified,  with  no 
stratification,  and  40  sites  have  been  selected  among  a  systematic  sample  base, 
with  proportionality  to  the  agricultural  activity.  The  FAO  Forest  Resources 
Assessment  1990,  following  recommendations  given  by  statisticians  (Czaplewski, 
1991)  is  using  a  sample  of  117  TM  sites  located  randomly  in  the  tropical  belt 
(FAO,  1993).  The  variables  to  be  estimated  are  the  forest  areas  and  the  change  of 
forest  cover  in  the  last  10  years.  The  sampling  intensity  is  10  %,  and  a  stratification 
has  been  performed  using  ecological  criteria.  The  sample  size  has  been  fixed  with 
an  objective  of  5%  of  the  standard  error.  The  deforestation  rate  was  estimated 
with  an  error  of  12.5  %,  which  can  be  explained  by  the  high  coefficient  of  variation 
of  deforestation.  The  FIRS  project  (Forest  Information  by  Remote  Sensing)  is 
another  activity  of  JRC,  Italy.  One  of  its  fundamental  activities  was  to  stratify 
Europe  into  115  homogeneous  strata  with  three  main  criteria  :  proportion  of  forest 
cover,  species  composition,  and  timber  volume.  The  strata  are  grouped  into  6 
statistical  regions,  where  223  sites  have  been  selected  with  unaligned  systematic 
sampling.  The  sampling  intensity  is  3  %  of  the  total  area,  or  10  %  of  the  forest 
lands.  The  TREES  project  (Tropical  Ecosystem  Environment  observation  by 
Satellites)  is  using  a  double  sampling  scheme  with  regression  between  AVHRR 
derived  forest  cover  proportion  as  the  auxiliary  variable  and  TM  derived  forest 
cover  proportion  as  the  target  variable.  The  sampling  unit  is  corresponding  to  a 
block  of  1  lxl  1  AVHRR  pixels.  A  total  of  1 121  units  have  been  selected,  each  TM 
scene  generating  more  thant  200  units  (cluster  sampling).  A  regression  is 
performed  between  the  forest  proportions  estimated  at  TM  and  AVHRR  levels. 

DESIGNING  AN  OPTIMUM  SCHEME 

Constraints 

The  constraints  in  designing  an  operational  sampling  scheme  are  related  to  the 
satellite  itself  (availability  of  data,  acquisition  delays,  repetitivity  for  multitemporal 
analysis...),  the  operationality  of  the  system  (fulfilement  of  the  requirements  and 
technical  feasibility),  efficiency  and  cost-effectiveness  (the  measurement  error 
induced  by  remote  sensing  techniques  must  be  compared  and  added  to  the 
sampling  error),  and  the  client  requirements  (terms  of  reference  and  required 
precision  at  which  level  of  confidence). 


378 


Definition  of  a  sound  sampling  scheme  using  H  RSI) 


At  a  very  first  stage,  it  is  necessary  to  identify  which  land  cover  parameters  are 
to  be  analysed.  The  sampling  scheme  will  depend  on  the  characteristics  of  the  land 
cover  type  under  investigation,  e.  g.  its  surface  proportion  in  the  entire  domain 
(study  area),  the  size  of  the  elementary  components  of  this  land  cover  type  (size  of 
agricultural  plots  for  instance),  its  spatial  distribution  (presence  or  absence  in  the 
different  strata  of  the  study  area),  its  occuring  frequency  (proportion  of  the  land 
cover  type  in  the  strata),  and  its  temporal  variation  (crops  cycle  for  instance) 
Based  on  high  resolution  satellite  images,  the  variable  to  estimate  with  satisfactory 
precision  can  be  the  proportion  of  the  land  cover  type,  e.g.  the  proportion  of 
agricultural  areas  or  the  forest  areas.  It  can  be  also  the  variation  of  proportion  with 
a  multitemporal  data  set.  The  a  priori  knowledge  of  the  domain  is  a  critical  issue 
for  setting  up  a  sound  sampling  scheme.  The  sample  must  be  as  representative  as 
possible  of  the  entire  population.  If  limited  knowledge  is  available,  multiphase 
sampling  is  recommended.  In  table  1,  sampling  schemes  are  proposed  with  respect 
to  the  characteristics  of  the  item. 


Table  1. -Designing  a  sampling  scheme  based  on  the  characteristics  of  the  item 
 (Cochran,  1981).  


charateristics  of  item 

type  of  sampling  needed 

1.  widespread  throughout  the  region, 
occuring  with  reasonable  frequency  in  all 
parts. 

a  general  survey  with  low  sampling  intensity. 

2.  widespread  throughout  the  region  but 
with  low  frequency. 

a  general  survey,  but  with  a  higher  sampling 
ratio. 

3.  occuring  with  reasonable  frequency  in 
most  parts  of  the  region,  but  with  more 
sporadic  distribution,  being  absent  in 
some  parts  and  highly  concentrated  in 
others. 

for  best  results,  a  stratified  sample  with 
different  intensities  in  different  parts  of  the 
region.  Can  sometimes  be  included  in  a 
general  survey  with  supplementary  sampling. 

4.  distribution  very  sporadic  or  concentrated 
in  a  small  part  of  the  region. 

not  suitable  for  a  general  survey.  Requires  a 
sample  geared  to  its  distribution. 

The  sampling  type  has  to  be  examined  when  using  satellite  images.  Indeed,  the 
measurements  are  not  made  on  a  single  point,  but  within  an  area  (SPOT  scene  for 
instance).  The  results  may  be  affected  to  the  scene  centre  (sample  of  points  taken 
from  an  infinity  of  points)  or  to  the  entire  area  (area  frame  sampling  with  a  sample 
taken  from  N  sites  covering  the  entire  region).  The  definition  of  the  sample  base  is 
an  important  step  of  the  sampling  scheme  :  it  is,  in  principle,  composed  of  all  units 
(points  or  frames)  constituting  the  entire  region.  When  using  SPOT  scenes  centres 
as  the  sample  base  on  very  large  region,  a  geometric  problem  is  raised  by  the 
narrowing  of  tracks  towards  the  pole,  which  induces  of  higher  density  of  units.  The 
sampling  scheme  should  take  into  consideration  this  distortion  by  selecting  the 


379 


sample. units  on  a  proportional  basis  (proportional  to  the  distance  between  two 
orbits  for  instance).  A  stratification  is  then  often  required,  mostly  when  the  a  priori 
proportion  of  the  item  is  known  with  sufficient  accuracy.  The  ideal  variate  for 
stratification  is  the  value  of  the  variable  to  estimate.  In  practice,  other  source  of 
information  is  used,  such  as  ancillary  data  (existing  maps  or  statistics).  The 
stratification  process  can  lead  to  different  types  of  strata  :  the  strata  in  which  the 
item  is  almost  absent  and  in  which  no  sample  will  be  taken,  and  the  strata  with 
different  proportion  of  the  item  (from  low  to  high  proportion).  It  is  sometimes 
recommended  to  delineate  special  strata  which  correspond  to  other  criteria 
(mountainous  areas,  swampy  zones... etc).  It  is  important  to  measure  the  total  area 
of  each  stratum,  and  to  verify  that  the  sum  is  giving  the  total  area  of  the 
inventoried  region.  The  number  of  strata  will  be  low  if  the  a  priori  knowledge  of 
the  population  is  limited.  Systematic  sampling  in  two  dimensions  represent  a  good 
area  sampling  frame  scheme.  An  alternative  is  the  unaligned  sampling  (Cochran, 
1981).  A  combination  of  systematic  sampling  (selecting  the  cluster)  and  random 
sampling  (selecting  the  unit  in  the  cluster)  may  also  give  good  results.  The  number 
of  sample  units  within  each  stratum  may  be  calculated  with  the  following  equation 
(Eq.  1)  : 

j 

with  n  as  the  total  sample  size,  nj  the  sample  size  in  stratum  j,  Sj  the  size  of  stratum 
j,  and  Varj  the  variance  of  stratum  j.  The  total  sample  size  n  is  determined  by  the 
required  overall  precision  and  the  available  budget.  For  optimizing  the  design  (cost 
effectiveness),  the  Lagrange  equation  can  be  used  to  solve  the  system.  In  stratified 
sampling,  the  overall  estimate  of  the  mean  value  is  weighted  by  the  strata  sizes 
(Eq.  2)  : 

Y  =  &   (2) 

N 

with  Nj  as  the  stratum  size,  Yj  the  estimated  mean  value  in  stratum  j,  and  N  the 
total  number  of  sample  units.  The  total  error  of  the  design  is  a  sum  of  the  sampling 
error  and  the  measurements  error.  When  the  measurement  error  is  important,  it  is 
not  worth  designing  a  complex  sampling  scheme  with  a  very  high  precision  since 
the  overall  precision  will  be  low. 

CONCLUSION 

In  sampling  statistics,  the  estimates  are  always  given  with  a  confidence  interval 
and  a  probability  level.  This  means  that  we  are  not  absolutely  sure  of  the  error  we 


380 


have  made  with  the  sample.  It  is  sometimes  difficult  to  decide  how  much  error 
should  be  tolerated  by  the  client  or  the  user.  Moreover,  starting  from  the  user 
requirements  and  means  (budget  allocated  for  the  survey),  and  looking  into  the 
characteristics  of  the  parameter  to  be  estimated,  it  should  be  pointed  out  that  there 
is  no  single  optimum  sampling  design.  In  our  problem  consisting  in  optimizing  the 
estimate  of  the  sample  size  and  the  location  of  the  sample  units,  it  can  be 
concluded  that  geostatistical  concepts  should  be  further  investigated  before  being 
applied  in  an  operational  project.  The  combination  of  several  standard  statistical 
principles  often  lead  to  satisfactory  results  (systematic  sampling  at  first  stage  and 
random  sampling  at  second  stage  for  instance).  The  use  of  standard  principles 
always  simplifies  the  design  where  the  error  can  be  easily  calculated.  In  operational 
projects,  the  use  of  a  satellite  reference  grid  (SPOT  or  TM)  seems  to  be  the  best 
way  for  building  up  the  sampling  base  (FAO/FRA  90  and  MARS  projects). 
However,  in  high  latitudes,  the  distorsion  of  the  grid  has  to  be  examined  in  order 
to  avoid  an  oversampling.  This  issue  should  be  addressed  by  global  survey  over 
large  areas,  such  as  the  agricultural  statistics  project  in  Russia  which  has  just  been 
launched  under  European  Union  support  (TACIS)  and  in  which  SCOT  Conseil  is 
providing  technical  assistance. 

REFERENCES 

Atkinson,  P.,  1994,  Testing  the  efficiency  of  sampling  strategies  with  simulated  remotely 

sensed  data,  in  SFPT,  n°137,  pl2. 
Brion,  P.,  et  Fournier,  P.,  1995,  Etude  Geostat-Maroc.  Partie  statistique.  Rapport  CNES 

travaux  INSEE,  SCEES. 
Chevrou  R.B.,  DERF  Montpellier,  Mars  1988,  Inventaire  Forestier  National,  Methodes  et 

Procedures. 

Cochran  W.  G.,  Mars  1981,  Sampling  techniques,  third  edition. 
Czaplewski  R.L.,  Analysis  of  alternative  sample  survey  designs,  FAO  1991. 
Dagnelie  P.,  1973,  Theorie  et  methodes  statistiques,  Volume  1,  Centre  de  Documentation 
de  Toulouse. 

Dagnelie  P.,  1975,  Theorie  et  methodes  statistiques,  Volume  2,  Centre  de  Documentation 
de  Toulouse. 

Dunn  R,  Harrison  A.R.,  1993,  Two-dimensional  systematic  sampling  of  land  use,  Appl. 

Statistics,  42,  N°  4,  pp.  585-601. 
Fitzpatrick-Lins  K.,  Mars  1981,  Comparison  of  sampling  procedures  and  data  analysis  for 

a  land-use  and  land-cover  map,  PERS  vol.  47,  N  3,  pp.  343-351. 
Food  and  Agriculture  Organization,  1993,  Forest  resources  assessment  1990:  tropical 

countries,  FAO  Forestry  Paper  1 12,  Rome,  61pp. 
Gallego  FJ  ,  P  Vossen,  JF  Dallemand,  V  Perdigao,  Sampling  plans  in  MERA  Project, 

MARS  project,  1995. 

Houllier  F.,  6  Jum   1986,  Echantillonnage  et  modelisation  de  la  dynarruque  des 
peuplements  forestiers/Application  au  cas  de  l'lnventaire  Forestier  National,  These. 


381 


Iverson  L.  R.,  Cook  E.A.,  Graham  R.  L.,  1989,  A  technique  for  extrapolating  and 
validating  forest  cover  across  large  regions  -  Calibrating  AVHRR  data  with  TM  data, 
International  Journal  of  Remote  Sensing,  Vol.  10,  n°ll,  p.  1805-1812. 

Kleinn  C,  Dees  M,  Pelz  D.R.,  1993,  Sampling  aspects  in  the  TREES  project  -  global 
inventory  of  tropical  forests,  final  report,  Freiburg  Universitat  Germany,  contract  for 
JRC  Ispra,  Italie. 

McGwire  K.,  Friedl  M.,  Estes  J.E.,  1993,  Spatial  structure,  sampling  design  and  scale  in 
remotely-sensed  imagery  of  a  California  savanna  woodland,  International  Journal  of 
Remote  Sensing,  Vol.  14,  n°ll,  p.  2137-2164. 

Nelson  R.  and  Holben  B.,  1986,  Identifying  deforestation  in  Brazil  using  multiresolution 
satellite  data,  International  Journal  of  Remote  Sensing,  vol.  7,  pp.  429  -  448. 

Pai'vinen  R.  and  Pitkanen  J.,  1992,  Calibrating  AVHRR  data  with  TM  data  for  tropical 
forest  cover  assessment,  IUFRO  S  4.02.05  Wacharakitti  international  workshop 
"Remote  sensing  and  permanent  plot  techniques  for  world  forest  monitoring",  Pattaya 
Thailand. 

Raffy  M.,  Lacaze  B.,  Rambal  S.  and  Winkel  T.,  August  1994,  Identifying  spatial  patterns 

of  Mediterranean  landscapes  from  geostatistical  analysis  of  remotely-sensed  data, 

International  Journal  of  Remote  Sensing,  Volume  15,  Number  12,  Special  issue: 

scaling  in  remote  sensing. 
Raffy  M.,  Puech  C,  August  1994,  Thresholds  of  homogeneity  in  targets  in  the  landscape. 

Relationship  with  remote  sensing,  International  Journal  of  Remote  Sensing,  Volume 

15,  Number  12,  Special  issue:  scaling  in  remote  sensing. 
SCOT  CONSEIL  &  GAF  for  JRC  EMAP  Unit,  1995,  FIRS  Project,  Regionalization  and 

stratification  of  European  forest  ecosystems,  final  report. 
SCOT  CONSEIL,  CCE,  1994,  Documentation  Action  IV,  Estimations  rapides  des 

superficies  et  des  productions  potentielles  au  niveau  europeen,  Volume  I  -  Document 

de  synthese. 

SCOT  CONSEIL,  CCE,  1994,  Documentation  Action  IV,  Estimations  rapides  des 
superficies  et  des  productions  potentielles  au  niveau  europeen,  Volume  III  - 
Methodologie. 

Smiatek  G.,  1995,  Sampling  thematic  mapper  imagery  for  land  use  data,  Remote  Sensing 
of  Environment,  52:116-121. 


382 


The  Use  of  AVHRR  Satellite  Imagery 
to  Monitor  Boreal  Ecosystem  Forest 

Fires 

Donald  R.  Cahoon,  Jr.1  and  Brian  J.  Stocks2 


Abstract  -  Forest  fires  are  an  integral  part  of  the  natural  forces  that  shape 
the  composition  and  evolution  of  the  boreal  forest.  These  fires,  often  very 
intense,  commonly  grow  to  sizes  large  enough  for  satellite  monitoring 
using  the  National  Oceanic  and  Atmospheric  Administration  (NOAA) 
Advanced  Very  High  Resolution  Radiometer  (AVHRR)  instrument.  The 
AVHRR  instrument  acquires  1-km  and  4-km  resolution  (at  nadir)  satellite 
imagery  in  five  spectral  bands.  Multispectral  methods  have  been 
developed,  using  the  AVHRR  imagery,  to  estimate  the  size  and  growth  rate 
of  burned  forest.  The  AVHRR  burned  area  estimates  have  been  evaluated 
and  the  error  determined  for  both  the  1-km  and  4-km  imagery.  Applying 
these  methods,  the  spatial  distribution  and  extent  of  forest  fires  has  been 
mapped  for  the  1987  and  1992  Russian  boreal  forest  fire  season.  Due  to  a 
strong  contrast  in  the  total  area  burned  in  each  of  the  2  years,  the  large 
interannual  variability  of  burned  forest  is  demonstrated. 


INTRODUCTION 

The  boreal  ecosystem  is  a  forested  belt,  mostly  comprised  of  conifer  trees,  that 
stretches  across  the  Northern  Hemisphere's  circumpolar  countries.  The  boreal 
forest  covers  over  12  million  km2  of  the  Earth's  surface  which  is  about  10%  of  the 
total  land  surface  area.  In  terms  of  areal  extent,  the  boreal  forest  ranks  about  fifth 
compared  to  all  terrestrial  biomes,  but  ranks  second  only  to  tropical  forests  in  total 
plant  mass  (Whittaker  and  Likens  1975).  Given  the  ecosystem's  size  and  total 
available  plant  mass  it  is  not  surprising  that  the  World's  boreal  forests  contain 
almost  40%  of  all  terrestrial  carbon  (Kasischke  et  al.  1995a).  This  means  that  the 
boreal  forest  must  play  an  extremely  important  role  in  the  global  carbon  budget 
and  it  is  important  to  characterize  any  disturbance  and  process  that  perturbs  the 
carbon  cycle  within  this  ecosystem.  Fire  is  one  such  disturbance. 

Because  of  the  enormous  geographical  scale  for  which  the  boreal  forest 
extends,  a  trade-off  between  satellite  resolution,  accuracy  of  estimating  the  acreage 
burned,  and  frequency  of  satellite  coverage  has  been  made.  In  reality,  there  are 
three  satellite  systems  that  could  potentially  be  used  to  monitor  forest  fires. 
Landsat  and  SPOT  quickly  come  to  mind  because  of  their  high  spatial  resolution, 
but  do  not  offer  the  repeat  coverage  needed  to  compensate  for  cloud  cover 


Research  Scientist,  NASAJLangley  Research  Center,  Hampton,  VA. 

Senior  Research  Scientist,  Canadian  Forest  Service,  Sault.  Sle.  Marie,  ONT,  Canada. 


383 


obscuration  of  the  land  surface.  Also,  the  amount  of  data  and  cost  of  enough 
imagery  to  cover  the  entire  boreal  region  would  be  unrealistic.  Rather,  we  have 
chosen  to  use  AVHRR  imagery  because  it  offers  a  reasonable  resolution  at  a  low 
cost,  and  it  provides  multiple  overpasses  each  day.  After  an  evaluation  of  the 
AVHRR  imagery,  we  have  found  that  estimates  of  the  surface  area  can  be  within  a 
few  percent.  This  error  is  considered  reasonable  for  the  large  geographical  scale 
we  are  evaluating  as  long  as  consistent  results  can  be  obtained. 

To  develop  consistent  surface  area  estimation  results,  we  have  developed  a 
methodology  for  processing  the  AVHRR  imagery.  The  following  sections  will 
describe  the  AVHRR  instrument  and  archived  imagery,  the  burned  area  estimation 
methodology  and  error  analysis,  and  describe  some  of  our  ongoing  research  on 
boreal  forest  fires. 

AVHRR  INSTRUMENT 

The  AVHRR  instrument  has  flown  on  several  National  Oceanic  and 
Atmospheric  Administration  (NO  A  A)  sun- synchronous  polar-orbiting 
meteorological  satellites  since  1978  and  provides  a  nearly  continuous  archived 
data  set.  These  satellites  have  flown  in  a  morning  and  afternoon  configuration 
which  provides  four  overpasses  each  day  in  the  low  to  mid-latitudes,  and  upwards 
toward  8  overpasses  in  the  higher  latitudes  of  the  boreal  forest.  The  AVHRR 
instrument  is  a  scanning  imager  which  scans  in  the  across  track  direction  and 
acquires  data  in  five  spectral  bands  (table  1).  The  swath  width  is  just  over  2800 
km  given  a  nominal  orbital  altitude  of  830  km  and  a  total  scan  angle  of  110.8°. 
Each  cross  track  scan  provides  2048  observations  (pixels)  with  an  instantaneous 
field-of-view  (IFOV)  of  approximately  1.4  mrad.  The  precise  IFOV  for  each 
channel  varies  slightly  and  is  included  in  table  1. 


Table  1.  -  AVHRR  channel  characteristics. 


Channel 

Wavelength 
(microns) 

Cross  track 
IFOV  (mrad) 

Along  track 
IFOV  (mrad) 

1 

0.58  -  0.68 

1.443 

1.433 

2 

0.725-  1.10 

1.433 

1.423 

3 

3.55  -  3.93 

1.423 

1.423 

4 

10.3  - 11.3 

1.280 

1.463 

5 

11.5-12.5 

1.320 

1.310 

Channel  selection  of  the  AVHRR  instrument  was  driven  by  operational 
meteorological  and  oceanographic  requirements.  However,  over  time  the  fire 
community  has  found  ways  to  apply  the  channel  selection  to  its  own  research.  In 
summary,  channel  1  and  channel  2  can  be  used  to  differentiate  smoke  from  clouds 


384 


(Chung  and  Le  1984).  Channel  2,  because  of  its  sensitivity  to  the  reflectance  from 
vegetated  surfaces,  provides  tremendous  contrast  between  burn  and  unburned 
areas  (Cahoon  et  al.  1992;  Cahoon  et  al.  1994;  Kasischke  et  al.  1995a).  Channel  3 
has  long  been  demonstrated  to  be  sensitive  to  hot  temperature  sources  (Matson  and 
Dozier  1981;  Scorer  1987)  and  is  useful  for  mapping  active  forest  fires  (Cahoon  et 
al.  1994;  Stocks  et  al.  1996).  Channels  4  and  5  are  both  similar  in  that  they  are  low 
in  contrast  when  calibrated,  but  provide  a  basis  for  distinguishing  between  burned 
areas  and  water  in  mid-afternoon  imagery  (Cahoon  et  al.  1994). 

The  imagery  is  archived  in  either  of  two  formats,  local  area  coverage  (LAC)  or 
global  area  coverage  (GAC).  The  LAC  format  is  the  full  1-km  (at  nadir) 
resolution  product  and  is  recorded  onboard  the  spacecraft  and  downlinked  to 
specific  sites.  The  AVHRR  1-km  data  are  also  continuously  broadcast  to  any 
ground-station  within  the  satellite's  horizon;  which  is  referred  to  as  high  resolution 
picture  transmission  format  (HRPT)  data.  In  the  archives  HRPT  and  LAC  data  are 
the  same.  Due  to  a  limited  amount  of  onboard  memory  for  recording  LAC  data,  a 
4/15  subsampled  product  is  continuously  recorded  for  broadcast  to  selected  ground 
stations.  This  format  is  the  global  area  coverage  (GAC)  and  is  frequently  referred 
to  as  the  4  km  imagery.  The  GAC  subsampling  scheme  is  shown  in  figure  1. 


Averaged  and  Saved 


"™"  ™" mm  [   ]  Discarded 

Figure  1.  -  GAC  subsampling  scheme.  Each  row  is  a  portion  of  a  scan  line. 
Each  box  is  one  image  pixel.  The  dark  box  represents  a  GAC  "4  km"  pixel. 

BURNED  AREA  ESTIMATION  METHODOLOGY 

The  surface  area  of  burn  scars  is  estimated  using  either  LAC  or  GAC  archived 
imagery.  In  a  previous  study,  Cahoon  et  al.  (1992)  demonstrated  that,  given  the 
large  fire  sizes  in  the  boreal  system  that  GAC  imagery  is  sufficient  for  estimating 
the  area.  In  the  following  section  we  will  examine  the  surface  area  estimation 
error  using  both  the  LAC  and  GAC  imagery.  This  section  will  lay  out  the 
methodology  through  which  we  process  the  AVHRR  imagery. 

Because  of  the  intermittent  availability  of  clear-sky  imagery,  which  negates  the 
possibility  of  observing  the  entire  study  region  at  one  time,  a  clear-sky  mosaic  is 
created.  From  this  mosaic,  all  burn  scars  can  be  observed  in  a  single  image, 
making  the  area-burned  estimate  simpler  and  alleviating  the  concerns  of  double 
counting  or  missing  any  burned  area.  The  development  of  the  clear-sky  mosaic  is 


385 


a  several-step  process  which  will  be  described  with  the  use  of  matrix  notation.  An 
individual  multispectral  (3-banded)  AVHRR  scene  is  represented  by  matrix  Dn  and 
the  mosaic  of  all  the  individual  AVHRR  scenes  is  matrix  M,  where 


dn. 

m12  •■ 

•  mij 

Dn  = 

d2l 

d22' 

■d2J 

•M  = 

m21 

m22  .. 

•  m2j 

di2- 

m„ 

mi2  .. 

(1) 


n  is  the  total  number  of  AVHRR  scenes;  i  and  j  define  the  number  of  elements,  d 
and  m,  in  the  matrices.  Each  pixel  in  the  AVHRR  image  is  corrected  for  brightness 
changes  due  to  the  solar  zenith  angle.  Using  the  NOAA-supplied  navigation  data, 
each  element  dy  is  the  result  of  resampling  the  original  NOAA  scene  into  a 
cylindrical-equidistant  map  projection.  Since  each  AVHRR  scene  is  resampled, 
for  every  Dn  the  elements  represent  the  same  geographical  region  and  are  a 
function  of  latitude  ((())  and  longitude  (X).  The  mosaic  image  (M)  is  the  same  size 
and  spatially  represents  the  same  geographical  area  as  the  AVHRR  scenes  Dn. 
Then  each  element  of  both  M  and  Dn  can  be  compared,  with  the  results  of  the 
comparison  composited  into  a  final  scene  (M').  The  comparison, 

M'u  =  MIN  (Dnij,Mij)  (2) 

tests  for  the  minimum  between  the  mosaicked  image  (M)  and  the  AVHRR  scene 
(Dn)  in  the  AVHRR  channel  2  band,  and  saves  the  elements  for  all  channels  of  the 
minimum-valued  image  in  the  new  mosaic  M'.  The  testing  is  iterative,  where  M' 
is  substituted  for  M,  and  continues  n  times  through  all  of  the  scenes.  Since  clouds 
are  highly  reflective  and  the  burned  area  is  of  low  reflectance  in  the  AVHRR 
channel  2,  the  burned  areas  are  retained  while  the  clouds  are  removed. 

In  estimating  the  total  regional  burned  area,  it  was  first  necessary  to  classify 
which  image  pixels  contained  burned  area.  An  unsupervised  minimum  distance 
classification  on  AVHRR  channels  1,  2,  and  4  was  than  used  iteratively  to  isolate 
the  scar  pixels  and  statistically  determine  the  edge  pixels  that  are  counted  as 
having  a  burn  scar.  The  iterative  process  continues  until  burn  scars,  that  have  been 
correlated  with  known  fire  activity  determined  from  the  AVHRR  imagery,  are 
isolated  and  falsely  classified  pixels  have  been  minimized. 

The  resulting  scar  map  is  subjected  to  another  processing  step  to  determine  the 
area  of  each  pixel.  Since  a  cylindrical-equidistant  projection  is  used  and  each  side 
of  a  pixel  is  parallel  to  both  meridians  and  parallels,  the  distance  along  each  pixel 
in  terms  of  degrees  of  latitude  and  longitude  is  easy  to  determine  as  well  as  the 
surface  area.  The  area  of  every  pixel  classified  as  a  scar  is  then  integrated  to  derive 
the  total  area  burned. 


386 


BURNED  AREA  ERROR  ANALYSIS 


An  extensive  analysis  has  been  conducted  to  evaluate  the  error  of  estimating 
the  bumed  surface  area  using  both  the  LAC  and  GAC  imagery.  The  details  are 
outlined  in  Cahoon  et  al.  (1992).  In  brief,  surface  targets  of  known  area  are 
delineated  in  both  the  LAC  and  GAC  imagery.  These  targets  are  a  series  of  lakes 
in  Canada  that  often  contain  several  islands,  sometimes  of  considerable  size,  and 
thus  are  not  homogeneous.  The  heterogeneous  nature  of  the  lake  targets  is  similar 
to  that  of  burn  scars  which  contain  unburned  areas  within  a  its  perimeter.  The 
results  from  this  analysis  are  in  figure  2. 

The  mean  surface  area  estimation  error  associated  with  using  LAC  imagery  is 
lower  than  that  of  GAC  imagery  across  the  entire  size  range,  even  though  for 
larger  targets  the  difference  is  practically  negligible.  This  comes  as  no  surprise 
since  we  expect  the  analysis  of  higher  resolution  imagery  to  yield  better  results 
when  working  with  small  areas.  However,  it  is  interesting  to  note  that  for  larger 
targets,  the  error  variance  for  the  GAC-derived  estimates  is  less  than  that  of  the 
LAC  analysis.  This  likely  has  a  lot  to  do  with  the  overlap  of  the  AVHRR  pixels 
(Cahoon  et  al.  1992).  The  overlap  is  considerable  for  the  1-km  data  and  the  edge 
of  a  target  is  defined  by  a  gradual  gradient,  leaving  room  for  interpretation  as  to 
where  to  define  the  target's  edge.  Much  of  the  surface  area  estimate  errors  are  due 
to  blurred  edges  of  targets.  At  least  for  the  GAC  imagery,  some  of  this  overlap  has 
been  eliminated  by  subsampling,  leaving  a  better  defined  edge  and  fewer  perimeter 
pixels  to  skew  the  analysis  one  way  or  another. 

Overall,  the  results  show  that  for  LAC  imagery  a  mean  error  of  about  2%  can 
be  expected,  and  that  90%  of  all  targets  will  fall  within  4%  error  for  targets  down 
to  100  km2.  For  the  GAC  imagery,  the  mean  error  is  about  2%  worse  than  that  of 
the  LAC  imagery  for  targets  that  are  100  km2,  and  drops  off  to  being  negligible 
when  evaluating  targets  that  are  10,000  km2  or  more  in  size.  Burned  area  estimate 
errors,  from  both  the  1988  Yellowstone  Fires  (Cahoon  et  al.  1992)  and  the  1987 
Great  Dragon  Fire  in  China  (Cahoon  et  al.  1994),  are  consistent  with  these  results. 

RESEARCH  APPLICATION:  RUSSIAN  BOREAL  FOREST  FIRES 

Forest  fires  have  been  a  major  disturbance  regime  affecting  the  world's  boreal 
forests  for  millennia.  Increased  development  within  the  world's  boreal  zone  has 
necessitated  fire  management  programs  designed  to  protect  human  interest  and 
forest  investment.  While  these  programs  have  been  largely  effective  in  reducing 
unwanted  fires  in  selected  regions,  forest  fires  continue  to  exert  enormous 
influence  in  the  world's  boreal  forests.  With  the  rising  concern  that  potential 
global  warming  will  lead  to  an  increase  of  fire  activity  within  the  boreal  forests 
(Flannigan  and  Van  Wagner  1991,  Wotton  and  Flannigan  1993,  Stocks  1993, 
Fosberg  et  al.  1996),  much  attention  as  recently  been  focused  on  the  important  role 
the  boreal  forest  plays  in  the  Earth's  carbon  budget  (Kasischke  et  al.  1995b),  and 
how  an  increase  in  fire  activity  will  impact  that  budget  and  potentially  provide 


387 


Area  (km2) 

Figure  2.  -  AVHRR-derived  surface  area  estimation  error  for  various  target 
sizes.  LAC  errors  shown  by  dashed  line;  GAC  errors  shown  by  solid  line.  The 
lower  line,  for  both  LAC  and  GAC,  represents  mean  error  and  the  upper  line 
shows  the  upper  error  bound  for  90%  of  all  targets. 

positive  feedback  to  global  warming  (Kurz  et  al.  1994).  Given  these  concerns  and 
the  need  for  a  quantitative  assessment  and  a  historical  perspective  of  boreal  fire 
activity,  we  have  begun  to  use  the  satellite  techniques  described  in  this  paper  to 
assess  the  total  area  burned  in  the  boreal  forest  during  the  last  15  years  for  which 
AVHRR  data  are  archived. 

Leading  into  this  multiyear  study,  as  test  cases,  we  have  completed  the  burned 
area  analysis  for  the  Russian  boreal  forest  for  the  years  1987  (Cahoon  et  al.  1994) 
and  1992  (Cahoon  et  al.  1996).  The  2  years  were  totally  dissimilar  in  terms  of  area 
burned,  with  1987  being  a  severe  fire  year  for  both  Russia  and  the  northern 
Chinese  provinces.  It  was  during  1987  that  one  of  the  largest  recorded  fires  in 
China's  history  occurred  (about  1.3  million  ha)  and  large  fires  throughout  eastern 
Russia  were  very  prevalent  (Cahoon  et  al.  1994).  The  1987  analysis  was 
conducted  using  over  200  GAC  images.  As  demonstrated  in  the  previous  section, 
the  burned  surface  area  estimation  using  GAC  imagery  is  not  much  greater  than 
6%  for  fire  sizes  down  to  100  km2  in  size.  Fortunately,  in  the  boreal  forest,  fire 
sizes  are  very  large  with  95%  of  the  area  burned  by  5%  of  the  fires.  The  estimate 
of  the  total  area  burned  during  the  1987  fire  season  is  14.446  million  ha. 

In  contrast,  even  though  there  was  widespread  fire  activity  in  1992,  the  total 
area  burned  was  1.5  million  hectares.  This  is  a  relatively  low  annual  total  of 
burned  acreage  which  was  concurrently  experienced  in  Alaska  and  Canada.  The 
1992  analysis  used  over  100  LAC  images  that  cover  the  region  from  30°  to  180° 
east  longitude.  The  boreal  region  was  split  into  30°  longitude  blocks.  Each  block 
was  processed  one  at  a  time  (Cahoon  et  al.  1996). 

There  was  an  obvious  similarity  between  1987  and  1992,  widespread  fire 
activity  is  primary  east  of  the  Yenisey  River  (about  90°N  latitude).  The  reason  for 
the  higher  amount  of  burned  area  to  the  east  is  likely  due  to  the  increase  in 
population  toward  the  west  and  the  low  marshy  areas  west  of  the  Yenisey  River. 
With  the  population  increase  to  the  west  comes  additional  fire  management 
pressures  and  better  access  to  fires  for  fire  suppression  purposes. 


388 


CONCLUSIONS 


With  limited  resources  and  the  constant  need  to  assess  spatial  changes  within 
our  natural  resource  regimes,  satellite  and  aircraft  remote  sensing  is  being  drawn 
upon  to  provide  spatially  dependant  data  over  various  scales.  The  need  to  monitor 
and  assess  the  post-burned  size  of  forest  fires  is  one  such  role  that  satellite-based 
observation  can  play.  Even  though  it  is  more  difficult  to  used  space-based 
resources  to  contribute  directly  to  fire  fighting,  it  is  not  out  of  the  question  to 
utilize  this  resource  to  monitor  fire  activity  and  map  post-burned  fires  over  large 
geographical  scales  and  remote  regions. 

The  remoteness  of  the  boreal  ecosystem,  a  region  where  considerable  travel 
and  fire  reconnaissance  is  accomplished  by  air,  almost  dictates  the  need  in  a  cost- 
cutting  era  to  assess  fire  activity  using  satellite  imagery.  The  remoteness, 
combined  with  the  current  interests  in  developing  a  consistent  fire  record 
throughout  the  entire  boreal  forest,  has  lead  us  to  develop  the  AVHRR  techniques 
used  in  our  two  trial  studies,  which  will  be  implemented  as  we  embark  on  a  15- 
year  analysis  of  the  interannual  variations  of  burned  forest.  Further  evaluations 
will  be  made  regarding  the  successfulness  of  detecting  and  estimating  the  size  of 
smaller  fires.  Further,  a  test  case  has  also  been  developed  which  demonstrates  the 
potential  use  of  the  AVHRR  imagery  for  monitoring  fire  growth  rates  (Stocks  et  al 
1996). 

To  conduct  this  research,  customized  computer  code  has  been  developed  over  a 
number  of  years  to  implement  our  approach.  Each  segment  of  our  software  system 
has  been  carefully  tested.  We  have  begun  to  evaluate  the  potential  of  commercial 
packages  for  registering  and  mapping  AVHRR  images.  In  one  prominent  GIS/ 
image  processing  package,  we  have  had  very  limited  success  with  registration  and 
would  not  trust  the  resulting  surface  area  estimates.  This  does  lead  to  our  concern 
that  any  commercial  tools  that  can  map  fires  and  estimate  the  size  of  burned  areas 
using  AVHRR  satellite  imagery  should  carefully  be  scrutinized  for  reliability. 

REFERENCES 

Cahoon,  Jr.,  D.  R.,  B.  J.  Stocks,  J.  S.  Levine,  W.  R.  Cofer  III,  and  C.  C.  Chung, 
1992,  Evaluation  of  a  technique  for  satellite-derived  estimation  of  biomass 
burning,  J.  Geophvs.  Res..  97(D4):  3805-3814. 

Cahoon,  Jr.,  D.  R.,  B.  J.  Stocks,  J.  S.  Levine,  W.  R.  Cofer  III,  and  J.  M.  Pierson, 
1994,  Satellite  analysis  of  the  severe  1987  forest  fires  in  northern  China  and 
southeastern  Siberia,  J.  Geophvs.  Res..  99(D9):  18,627-18,638. 

Cahoon,  Jr.,  D.  R.,  B.  J.  Stocks,  J.  S.  Levine,  W.  R.  Cofer,  and  J.  A.  Barber,  1996, 
Monitoring  the  1992  Forest  Fires  in  the  Boreal  Ecosystem  using  NO  A  A 
AVHRR  Satellite  Imagery:  Biomass  Burning  and  Global  Change,  MIT  Press, 
Cambridge,  MA.  (in  press) 

Chung,  Y.  S.  and  H.  V.  Le,  1984,  Detection  of  forest-fire  smoke  plumes  by  satellite 
imagery,  Atmos.  Environ..  18(10):  2143-2151. 


389 


Flannigan,  M.  D.,  and  C.  E.  Van  Wagner,  1991,  Climate  change  and  wildfire  in 

Canada,  Can.  J.  For.  Res..  21:  66-72. 
Fosberg,  M.  A.,  B.  J.  Stocks,  and  T.  J.  Lynham,  1996,  Risk  analysis  in  strategic 

planning:  fire  and  climate  change  in  the  boreal  forest:  Fire  in  Ecosystems  of 

Boreal  Eurasia,  Blackwell  Press,  Oxford,  UK.  (in  press) 
Kasischke,  E.  S.,  N.  H.  F.  French,  L.  L.  Bourgeau-Chavez,  and  N.  L.  Christensen, 

Jr.,  1995,  Estimating  release  of  carbon  from  1990  and  1991  forest  fires  in 

Alaska,  J.  Geophvs.  Res..  100:  2941-2951. 
Kasischke,  E.  S.,  N.  L.  Christensen,  and  B.  J.  Stocks,  1995,  Fire,  global  warming, 

and  the  carbon  balance  of  boreal  forests,  Ecol.  Appl..  5(2):  437-451. 
Kurz,  W.  A.,  M.  J.  Apps,  B.  J.  Stocks,  and  W.  J.  A.  Volney,  1994,  Global  climate 

change:  disturbance  regimes  and  biospheric  feedbacks  of  temperate  and  boreal 

forests:  Biotic  Feedbacks  in  the  Global  Climate  System:  Will  the  Warming 

Speed  the  Warming?:  Oxford  Univ.  Press,  Oxford,  UK. 
Matson,  M.,  and  J.  Dozier,  1981,  Identification  of  subresolution  high  temperature 

sources  using  a  thermal  IR  sensor,  Phot.  Eng.  Remote  Sensing,  47(9):  1311- 

1318. 

Scorer,  R.  S.,  1987,  Hot  Spots  and  Plumes:  Observation  by  Meteorological 

Satellite,  Atmos.  Environ..  21  (6):  1427-1435. 
Stocks,  B.  J.,  1993,  Global  warming  and  forest  fires  in  Canada,  For.  Chron..  69: 

290-293. 

Stocks,  B.  J.,  D.  R.  Cahoon,  Jr.,  W.  R.  Cofer  III,  and  J.  S.  Levine,  1996, 
Monitoring  large  scale  fire  behavior  in  northeastern  Siberia  using  NOAA- 
AVHRR  satellite  imagery,  Biomass  Burning  and  Global  Change,  MIT  Press, 
Cambridge,  MA.  (in  press) 

Whittaker,  R.  H.,  and  G.  E.  Likens,  1975,  Primary  Production:  The  biosphere  and 
man:  Primary  Productivity  of  the  biosphere,  Ecological  Studies  14,  Springer 
Verlag,  New  York. 

Wotton,  B.  M.,  and  M.  D.  Flannigan,  1993,  Length  of  fire  season  in  a  changing 
climate,  For.  Chron..  69:  187-192. 

BIOGRAPICAL  SKETCH 

Donald  R.  Cahoon,  Jr.  is  a  research  scientist  with  the  National  Aeronautics  and 
Space  Administration,  Langley  Research  Center,  Hampton,  Virginia.  He  has 
worked  as  an  analyst  of  satellite,  aircraft,  and  meteorological  data  sets  for  over  1 1 
years.  During  the  last  7  years  he  has  specialized  in  satellite  remote  sensing.  His 
primary  applications  have  been  forest  fire  monitoring  and  surface  radiation  budget 
studies. 

Brian  J.  Stocks  is  a  senior  forest  fire  scientist  with  the  Canadian  Forest  Service 
in  Sault  Ste.  Marie,  Ontario.  Over  the  past  28  years  he  has  specialized  in  the 
development  of  forest  fire  danger  rating  and  fire  behavior  prediction  systems,  and 
is  currently  involved  in  the  investigation  of  cause-and-effect  relationships  between 
global  change  and  forest  fires. 


390 


Large  Area  Forest  Cover  Assessment: 
Effects  of  Misregistration  in  a  Double 
Sampling  Approach  with  Coarse  and 
High  Resolution  Satellite  Images 

Christoph  Kleinn1,  Berthold  Traub1,  Matthias  Dees1 

Abstract  When  combining  satellite  data  of  considerable  different  spatial 
resolution  the  effect  of  misregistration  is  one  of  the  technical  issues  to  be 
addressed.  This  simulation  study  is  based  upon  subcontract  research  carried 
out  in  the  framework  of  tne  TREES  project  (Tropical  Ecosystem 
Environment  Observation  by  Satellites,  Joint  Research  Centre,  Ispra,  Italy). 
Overall  objective  is  a  global  tropical  forest  cover  estimate  and  the  production 
of  a  forest  map.  In  a  first  phase  of  the  inventory  approach  a  complete 
coverage  by  the  coarsly  resolving  NOAA  AVHRR  satellite  images  is 
provided.  To  improve  the  forest  area  estimates  derived  from  this  image  set  a 
sample  of  Landsat-TM  scenes  was  selected  in  a  second  phase:  In 
corresponding  frames/blocks  of  the  same  geographical  area  forest  cover 
percent  in  AVHRR  and  TM  was  recorded.  These  pairs  of  values  formed  the 
mput  variables  for  a  calibration  regression. 

To  assume  perfect  geographic  registration  is  certainly  not  realistic.  Some 
effects  of  misregistration  Detween  the  coarse  and  high  resolution  image 
frames/pixel  blocks  on  the  resulting  regression  are  studied  in  this  paper  in  tne 
form  of  a  simulation  study.  The  effects  can  be  considerable,  particularly  when 
small  blocks  of  coarse  resolution  pixels  are  to  be  registered.  

1.      INTRODUCTION  AND  OBJECTIVES 

Global  monitoring  of  forest  cover  is  a  current  issue  in  the  context  of  the 
destruction  of  tropical  forests  and  in  the  discussion  about  global  climatic  changes. 
Forest  cover  monitoring  systems  are  to  provide  sound  information  on  state  and 
changes  in  forest  cover.  Satellite  remote  sensing  plays  an  important  role  there: 
NOAA  AVHRR  has  proved  to  have  favourable  characteristics  to  discriminate 
vegetation  from  other  land  cover  and  is  frequently  used  in  environmental 
monitoring  (Ehrlich  et  al.  1994).  LANDSAT  TM  has  in  the  1980  s  developed  to 
be  the  standard  high  resolution  sensor  in  many  forestry  applications.  These  two 
systems  have  quite  different  characteristics  with  respect  to  temporal,  spatial  and 
spectral  resolution,  in  detail  described  in  standard  textbooks. 

The  high  temporal  frequency  and  low  cost  of  the  coarsely  resolving  AVHRR 
(about  1  km  x  1  km)  is  well  suited  to  provide  in  a  first  phase  a  more  or  less 
complete  coverage  of  the  region  of  interest.  This  results  in  an  equally  coarse' 
estimation  of  forest  cover.  The  estimate  might  then  in  a  second  phase  be  improved 
through  a  sample  of  the  much  higher  resolving  LANDSAT  TM,  which  is  much 
more  detailed  but  also  much  more  expensive.  Regression  technique  is  used  to 


1  Abteilung  fur  Forstliche  Biometrie,  Universitat  Freiburg,  Werderring  6,  D-79085  Freiburg, 
Germany  (kleinn@forst.uni-freiburg.de) 

2  The  acronym  fcp  is  used  throughout  the  paper  for  forest  cover  percent 


391 


predict  true'  forest  cover  from  the  coarse  information  delivered  by  AVHRR. 
True'  forest  cover  is  assumed  to  be  represented  by  the  TM  forest  cover.  In  terms 
of  sampling  theory  this  is  a  double  sampling  for  regression  approach  (Cochran 
1977).  From  an  image  interpretation  and  classification  point  of  view  the  resulting 
regression  can  be  regarded  as  calibration  function. 

To  calculate  the  regression  it  is  necessary  to  obtain  geographically  matching 
pairs  of  map  frames,  one  of  the  coarse  resolution  map  and  one  of  the  high  reso- 
lution map.  This  procedure  is  also  called  coregistration.  Two  approaches  how  to 
calculate  the  regression  are  discussed  in  the  literature: 

(1)  Coregistration  of  forest  percentages  from  both  data  sources;  an  example  is 
found  in  Nelson  (1989).  This  approach  is  also  used  in  a  global  tropical  forest 
assessment  in  the  TREES  project,  a  general  description  of  which  can  be  found  in 
Malingreau  (1993).  Units  to  be  coregistered  can  be  areas  corresponding  to  the  size 
of  one  single  coarse  resolution  pixel,  with  only  to  values  possible:0  and  1,  or 
blocks  of  pixels. 

(2)  Coregistration  of  forest  percentages  in  the  high  resolution  image  with 
spectral  values  of  coarse  resolution  image.  Examples  are  found  in  Iverson  et  al. 
(1989),  Paivinen  and  Pitkanen  (1992)  and  Zhu  and  Evans  (1992).  In  this  approach 
the  units  to  be  coregistered  are  areas  of  the  size  of  one  single  coarse  resolution 
pixel. 

Like  all  types  of  measurements  coregistration  is  subject  to  error,  too:  The 
problem  of  misregistration  arises.  This  paper  investigates  in  the  form  of  a 
simulation  study  possible  effects  of  misregistration  onto  the  calibration  procedure. 
It  is  an  "extension  of  a  former  study  presented  in  Kleinn  et  al.  (1995).  Approach  (1) 
is  pursued  here,  using  blocks  of  pixels  as  registration  unit. 

2.      GENERAL  DESCRIPTION  OF  MISREGISTRATION 

Not  many  publications  yet  deal  with  the  impact  of  misregistration  in  detail.  One 
of  the  few  articles  focuses  on  change  detection:  Townshend  et  al.  (1992)  state  that 
change  assessments  using  satellite  images  of  two  different  taking  dates  are  affected 
by  the  level  of  registration,  depending,  of  course,  on  several  factors.  They  finally 
state  that  "high  levels  of  registration  must  be  achieved  by  operational  monitoring 
systems  if  there  is  to  be  reliable  monitoring  of  global  change". 

Assuming  perfect  classification  procedures,  the  two  data  sources  under 
consideration  here  (coarse  and  high  resolution)  would  yield  identical  forest  cover 
estimates.  A  simple  linear  regression  between  the  data  pairs  would  result  in  the  one 
to  one  line.  In  reality,  there  are  more  or  less  significant  deviations:  Sensors  of 
different  spectral  and  spatial  resolution  'see'  things  in  a  different  way.  Additionally 
geometric  inaccuracies  lead  to  misregistration:  The  block  of  high  resolution  pixels 
is  not  exactly  matching  the  corresponding  block  of  coarse  resolution  pixels. 

Perfect  registration  would  mean  that  (1)  the  centres  of  the  two  blocks  match 
and  that  (2)  their  shape  and  (3)  their  alignment  are  the  same.  In  Figure  1  these 
factors  are  depicted  schematically.  Misregistration  means  that  not  only  the  high 
resolution  pixels  in  the  true  matching  block  have  a  chance  to  be  registered,  but  also 
the  pixels  around  them.  Would  the  amount  and  type  of  misregistration  be  known, 
then  one  could  calculate  the  probability  of  a  pixel  to  be  included  in  the  registration 
process  of  one  specific  block.  For  perfect  coregistration  this  probability  would  be  1 


392 


for  the  pixels  in  the  matching  block,  0  elsewhere.  In  the  presence  of  misregistration 
the  probability  is  high  in  the  matching  block  but  greater  than  0  in  a  certain  area 
around.it.  It  is  0  only  beyond  the  maximum  misregistration  distance.  Under  the 
assumption  that  misregistration  is  an  isotrope  process  this  area  is  drawn  in  Figure  1 
as  a  circular  envelope. 


Figure  1:  Large  circle:  Area  in  which 
the  registration  probability  of  the 
pixels  of  the  high  resolution  images  is 
greater  than  0.  Small  circle:  Area  of 
pixels  having  an  inclusion  probability 
greater  0  for  a  'rotating'  displaced 
block  with  fixed  distance  d  and  shift 
angle  a. 


Assumed  that  the  actual  registration  of  individual  sample  blocks  in  two  different 
images  is  a  random  process,  misregistration  leads  to  a  bias  when  estimating  fcp: 
The  expectation  of  the  attribute  (forest  cover)  measured  in  the  high  resolution 
image  is  not  the  value  of  the  true  matching  block.  This  adds  another  source  of 
variability  (error)  into  the  target  regression  between  forest  cover  estimates  of  high 
and  coarse  resolution  pixel  blocks. 

3.  MATERIAL 

Computer  generated  forest/non-forest  (1/0)  maps  of  size  9000x9000  pixels 
were  used  to  investigate  the  effects  of  misregistration.  A  set  of  'homogeneous' 
maps  was  created  by  randomly  locating  clusters  of  dots.  These  dots  have  diameters 
3,  30  and  100  pixels  for  the  map  sets  la-le,  2a-2e  and  3a-3e,  respectively.  The 
total  fcp  goes  from  10%  (in  maps  la,  2a,  3a)  over  30%,  50%,  70%  to  90%  (in 
maps  le-3e).  Three  inhomogeneous  maps  (maps  4  to  6)  consist  of  regions  with 
different  structures.  Examples  of  these  maps  at  original,  high  resolution  are  given 
in  Figure  2,  left  hand  side. 

The  high  resolution  maps  were  gradually  degraded  to  produce  coarse  resolution 
maps.  Square  pixel  blocks  of  size  n  x  n  pixels  of  the  original  image  were  collapsed 
to  one  new  coarse  resolution  pixel,  with  n  taking  on  the  values  10,  25  and  50. 
Depending  on  the  forest  cover  within  the  n  x  n  high  resolution  pixels  the  attribute 
forest  or  non-forest  was  assigned  to  the  new  coarse  resolution  pixel.  If  fcp 
exceeded  30%,  the  attribute  forest'  was  assigned.  In  Figure  2,  right  hand  side, 
results  of  the  degradation  are  shown.  For  the  degraded  maps  only  the  'core-region' 
of  8000x8000  pixels  is  shown  in  Figure  2,  the  region  to  which  the  analysis  is 
limited.  This  allows  for  a  buffer  frame  surrounding  the  analysed  region,  thus 
facilitating  the  treatment  of  edge  effects  in  the  simulations. 


393 


Map  4  (44.4%  cover):  Original  resolution  Map  4:  Degraded  (level  50,  see  text) 


Figure  2:  Sample  maps  as  used  in  the  simulation  study.  Left  hand  side:  Original  maps  of 
size  9000x9000  pixels.  For  the  degraded  maps  (right  hand  side)  only  the  core-region 
corresponding  to  8000x8000  pixels  in  original  resolution  is  depicted. 


394 


Table  1:  Forest'  cover  percent  in  the  original  and  degraded  maps  used  in  this  study. 
The  fcp  -values  of  the  maps  shown  in  Figure  2  are  framed. 


Map 

Original 

Degraded 

resolution 

10  x  10 

25  x25 

50  x  50 

forest  cover  percent  of  the  tot 

al  map 

Map 

10.0 

1  ~\  6 

I  J  .  VJ 

12.9  5.4 

1  K 
1  D 

30.0 

40.1  40.1 

lc 

50.0 

63.1 

77.3 

88.8 

Id 

70.0 

83.8 

95.5 

99.5 

le 

90.0 

97.8 

99.9 

100.0 

Map 

2a 

10.0 

8.5 

9.0 

9.8 

2b 

30.0 

32.0 

33.9 

36.5 

2c 

50.0 

52.4 

54.5 

57.0 

2d 

70.0 

72.1 

74.7 

78.3 

2e 

90.0 

91.2 

93.1 

95.2 

Map 

3a 

10.0 

10.4 

10.7 

11.0 

3b 

30.0 

30.2 

30.8 

31.6 

3c 

50.0 

51.0 

51.6 

53.1 

3d 

70.0 

70.6 

71.5 

72.9 

3e 

90.0 

90.6 

91.1 

92.0 

Map 

4  44.4 

48.1 

50.1 

51.5 

Map 

5 

32.6 

45.3 

51.7 

57.5 

Mar 

6 

34.1 

35.4 

37.8 

40.0 

The  image  degradation  technique  leads  to  a  change  in  total  forest  cover  as 
illustrated  in  Table  1 .  It  is  a  very  simple  technique,  which  probably  does  not  mirror 
very  realistically  the  properties  of  two  real  map  sets.  But  it  was  felt  that  this  was 
sufficiently  realistic  to  investigate  the  general  properties  of  misregistration.  With 
real  data  sets  the  differences  in  fcp  values  would  be  due  to  differences  in  spatial 
resolution,  taking  date,  atmospheric  conditions  etc.  and,  of  course,  differences  in 
image  interpretation  and  classification. 

4.  METHODS 

When  modeling  misregistration  one  has  to  make  assumptions  on  the  distance 
and/or  direction  distribution  of  the  deviations  relative  to  the  true  location.  The 
registration  process  is  modelled  as  purely  random  process  here.  The  factors  which 
were  included  in  this  simulation  and  which  were  subject  to  variation  at  several 
levels  are  listed  in  Table  2.  Non-matching  shapes  of  the  two  blocks  to  be  are  not 
taken  into  respect  in  this  study.  For  each  combination  of  the  listed  factor  levels  a 
systematic  grid  of  coarse  resolution  sample  blocks  was  superimposed  onto  the 
original,  high  resolution  maps.  For  all  blocks  at  first  the  fcp  of  the  true  matching 
high  resolution  block  was  determined.  Then  the  high  resolution  block  was 
displaced  100  times  according  to  the  probability  assumptions  given  in  Table  2.  For 
each  block,  mean  and  standard  deviation  of  the  fcp  values  resulting  from  the  100 
replications  were  recorded. 

5.  RESULTS 

Statistics  of  the  differences  between  truly  matching  pixel  block  and  the  mean 
(expectation)  of  the  100  simulated  misregistered  pixel  blocks:  In  Figure  1  it  was 
illustrated  that  a  bias  is  suspected  when  estimating  the  fcp  as  an  expectation  of 


395 


misregistered  pixel  blocks.  In  some  results  are  given  for  the  maps  shown  in  Figure 
2,  for  degradation  level  50.  The  bias  (mean  of  the  differences  for  all  blocks 
analysed  througout  the  whole  map)  is  negligible  for  all  maps  investigated. 


Table  2:  List  of  factors  included  in  the  model  study 


Factor  Description  and factor  levels 


Area  structure  Artificially  generated  maps:  Some  produced  with  a  homogeneous  genera- 

tion process,  some  with  several  processes  overlaid  (heterogeneous) 

Total  forest  cover  Total  forest  cover  is  10%,  30%,  50%,  70%  and  90%  for  the  homogeneous 

maps  and  32%  to  44%  for  the  heterogeneous  ones 

Spatial  resolution  of  coarse    Side  length  of  the  square  pixels  of  the  coarse  resolution  maps  correspond 
resolution  image  to  10,  25  and  50  pixels  of  the  high  resolution  image 

Forest/non-forest  rule  to  be    Minimum  crown'  cover  percent  in  the  degraded  image  for  a  pixel  to  be 
used  for  image  degradation    assigned  to  the  class  'forest'  is  fixed  to  be  30% 

Block  size  Coarse  resolution  square  pixel  blocks  of  2,  5,  10,  20  pixels  side  length 

Parameters  of  Misregistration  described  in  Figure  1 .  Maximum  distance  is  4  pixels  of 

misregistration  coarse  resolution  following  a  linearly  decreasing  pdf  (mean  distance  = 

4/3  =  1.333  pixels,  standard  deviation  =  Vl6/  18  =  0.943  pixels).  Isotropy 
is  assumed.  The  coregistered  blocks  may  rotate  up  to  ±10  degrees,  follow- 
ing a  triangular  probability  density  function  symmetric  around  0  degrees. 

For  the  single  sample  blocks  -  i.e.  not  using  the  mean  of  the  100  replications  per 
block  -  the  difference  between  truly  matching  pixel  block  and  misregistered  pixel 
bock  can  be  considerable.  This  can  be  seen  with  the  range  and  standard  deviation 
of  the  differences  as  given  in  Table  3:  For  Map  3d  and  block  size  2  x  2  the 
standard  deviation  is  about  10%  fcp,  having  a  range  of  differences  from  -41%  to 
+38%,  meaning  that  one  has  to  be  aware  of  large  variation  when  making  blockwise 
evaluations.  For  a  block  size  of  10  x  10  or  even  20  x  20  this  variation  is  much  less. 

This  variability  of  the  differences  is  also  a  function  of  total  forest  cover,  which 
can  be  observed  when  analysing  for  example  the  sequence  of  maps  la  to  le.  These 
maps  were  generated  with  the  same  algorithm  of  random  clustering  of  dots  of  a 
diameter  of  3  pixels.  The  difference  is  that  Map  la  was  only  filled  up  to  10%,  Map 
lb  up  to  30%  and  so  on.  Map  le  has  a  90%  cover.  In  Figure  3  the  standard 
deviation  of  the  differences  is  shown  over  total  forest  cover:  The  highest  variability 
is  thus  -reached  with  the  intermediate  fcp  of  around  50%.  With  a  forest  cover  of 
10%  or  90%  variability  is  much  less.  This  relationship  is  a  quite  obvious  one  when 
analysing  the  extreme  cases:  In  the  complete  presence  (or  absence)  of  forest, 
misregistration  has  no  effect  at  all. 

Calibration  regression:  The  regression  is  to  predict  blockwise  the  fcp  estimate 
of  the  coarse  resoution  pixel  blocks.  Here,  simple  linear  regression  is  used.  Though 
it  would  be  justiied  for  practical  reasons  the  intercept  was  not  suppressed.  Given 
the  situaion  illustrated  in  Table  1  it  is  clear  that  one  gets  a  slope  coefficient 
different  from  1  (in  most  cases  smaller  than  1),  when  predicting  fcp  of  the  high 
resoluion  map  (dependent  variable)  using  the  fcp  of  the  coarse  resolution  map  as 
independent  variable.  This  holds  even  for  the  situation  of  perfect  coregistration. 

In  analysing  the  regression  results  we  are  interested  in  two  relationships: 
Regression  1  under  perfect  coregistration  and  Regression  2  in  the  presence  of 


396 


misregistration.  In  Table  4  general  characteristics  of  the  regressions  are  listed  for 
the  three  sample  maps  addressed  throughout  this  paper.  Given  the  same  map  and 
the  same  degradation  level  there  are  considerable  differences  between  the 
coefficients  of  Regression  1  and  Regression  2  for  the  small  block  size  2,  but  the 
differences  level  out  quite  markedly  for  the  largest  block  size  20  Intercept  and 
slope  coefficients  of  the  Regressions  1  and  2  are  quite  close  for  this  large  block 
size.  As  Regression  1  is  the  true'  regression  under  perfect  coregistration  one  sees 
that  small  block  sizes  lead  to  an  incorrect  calibration  function  and  that  only  for 
very  large  block  sizes  Regression  2  approaches  the  coefficients  of  the  true 
Regression  1. 


Std 


#6.2-  '  *:4"  •  -*.3 


•>  6 


•2  3 


Figure  3:  Standard  deviation  of  differences 
between  fcp  of  truly  matching  pixel  block 
and  the  mean  of  the  100  misregistered  pixel 
blocks  over  fcp 


0  20         40         60         80  100 

Total  forest  cover  (fcp) 


Table  3:  Descriptive  statistics  of  the  differences  between  the  fcp  of  the  true  matching 
high  resolution  block  and  the  fcp  results  of  the  100  simulated  misregistered  blocks 
(degradation  level  50): 


Block  size 

Mean 

Std  Dev 

Minimum 

Maximum 

Map 

2 

0.000246 

0.0356 

-0.0931 

0.1679 

la 

5 

0.000281 

0.0103 

-0.0326 

0.0442 

10 

0.000255 

0.0036 

-0.0098 

0.0096 

20 

0.000256 

0.0011 

-0.0029 

0.0031 

Map 

2 

0.003002 

0.1017 

-0.4104 

0.3791 

3d 

5 

0.002945 

0.0429 

-0.1661 

0.1418 

10 

0.002647 

0.0156 

-0.0417 

0.0568 

20 

0.003288 

0.0061 

-0.0149 

0.0177 

Map 

2 

0.001279 

0.0821 

-0.4125 

0.4526 

4 

5 

0.001563 

0.0195 

-0.1277 

0.0690 

10 

0.001645 

0.0080 

-0.0338 

0.0274 

20 

0.000914 

0.0038 

-0.0111 

0.0125 

Differences  between  the  coefficients  of  Regression  1  and  Regression  2  are  most 
clear  for  map  set  1,  consisting  of  very  small  forest  patches'  compared  to  the 
spatial  resolution  of  the  degraded  images.  This  holds  -  as  mentioned  in  the 
preceeding  paragraph  -  particularly  for  small  block  sizes  and  low  total  forest  cover. 
For  larger  patches  and  higher  total  fcp  the  effect  is  much  smaller,  though  still  there 


397 


Table  4:  Basic  statistics  of  the  simple  linear  calibration  regression  for  three  sample  maps 
(rmse=root  mean  square  error,  r2=coefficient  of  determination). 


Map  la  (10%  cover  in  original  map) 


Deg 
lev. 

bi: 

size 

Regression  1 

Regression  2 

within 
block  std 

rmse 

r2 

Interc 

Slope 

rmse 

r2 

Interc. 

Slope 

10 

2 

0.0509 

0.9177 

0.0093 

0.7028 

0.0569 

0.7789 

0.0443 

0.4413 

0.1058 

5 

0.0201 

0.9618 

0.0042 

0.7404 

0.0246 

0.9283 

0.0164 

0.6483 

0.0406 

10 

0.0099 

0.9798 

0.0023 

0.7546 

0.0123 

0.9650 

0.0086 

0.7055 

0.0207 

20 

0.0049 

0.9873 

0.0014 

0.7611 

0.0057 

0.9816 

0.0037 

0.7416 

0.0081 

25 

2 

0.0507 

0.7581 

0.0451 

0.4640 

0.0447 

0.4970 

0.0750 

0.2296 

0.0769 

5 

0.0234 

0.8497 

0.0339 

0.5511 

0.0232 

0.7929 

0.0466 

0.4503 

0.0302 

10 

0.0120 

0.8892 

0.0291 

0.588 

0.0121 

0.8711 

0.0344 

0.5445 

0.0118 

20 

0.0057 

0.9091 

0.0286 

0.5924 

0.0058 

0.8980 

0.0315 

0.5675 

0.0047 

50 

2 

0.0507 

0.4728 

0.0848 

0.3735 

0.0338 

0.3682 

0.0938 

0.2006 

0.0531 

5 

0.0238 

0.5631 

0.0792 

0.4768 

0.0200 

0.5360 

0.0842 

0.3790 

0.0195 

10 

0.0113 

0.6388 

0.0777 

0.5051 

0.0107 

0.6099 

0.0805 

0.4486 

0.0078 

20 

0.0061 

0.7116 

0.0761 

0.5348 

0.0059 

0.7062 

0.0770 

0.5137 

0.0026 

Map  3d  (70%  fcp  in  original  map) 

Deg 
lev. 

Bl. 

size 

Regression  1 

Regression  2 

within 
block  std 

rmse 

r2 

Interc. 

Slope 

rmse 

r2 

Interc. 

Slope 

10 

2 

0.0331 

0.9945 

-0.0042 

0.9978 

0.0538 

0.9845 

0.0192 

0.9592 

0.0485 

5 

0.0165 

0.9985 

-0.0078 

1.0030 

0.0229 

0.9971 

0.0020 

0.9854 

0.0343 

10 

0.0101 

0.9994 

-0.0085 

1.0040 

0.0147 

0.9986 

-0.0014 

0.9905 

0.0345 

20 

0.0059 

0.9997 

-0.0088 

1.0044 

0.0064 

0.9997 

-0.0064 

0.9976 

0.0235 

25 

2 

0.0513 

0.9857 

-0.0107 

0.9949 

0.0757 

0.9641 

0.0445 

0.9143 

0.0972 

5 

0.0247 

0.9960 

-0.0192 

1.0067 

0.0346 

0.9915 

0.0087 

0.9642 

0.0749 

10 

0.0139 

0.9983 

-0.0210 

1.0093 

0.0188 

0.9967 

-0.0076 

0.9875 

0.0525 

20 

0.0075 

0.9991 

-0.0197 

1.0074 

0.0089 

0.9987 

-0.0135 

0.9954 

0.0299 

50 

2 

0.0690 

0.9707 

-0.0197 

0.9875 

0.1006 

0.9168 

0.0923 

0.8297 

0.1603 

5 

0.0325 

0.9905 

-0.0375 

1.0119 

0.0444 

0.9792 

0.0198 

0.9293 

0.1079 

10 

0.0170 

0.9953 

-0.0382 

1.0128 

0.0205 

0.9926 

-0.0116 

0.9728 

0.0581 

20 

0.0090 

0.9966 

-0.0410 

1.0168 

0.0108 

0.9950 

-0.0347 

1.0035 

0.0267 

Map 

4  (44%  fcp  in  original  map) 

Deg 
lev 

Bl. 

size 

Regression  1 

Regression  2 

within 
block  std 

rmse 

r2 

Interc 

Slope 

rmse 

r2 

Interc. 

Slope 

10 

2 

0.0625 

0.9817 

-0.0067 

0.9734 

0.0938 

0.9540 

0.0227 

0.9080 

0.0915 

5 

0.0324 

0.9943 

-0.0147 

0.9899 

0.0402 

0.9909 

-0.0056 

0.9677 

0.0523 

10 

0.0210 

0.9973 

-0.0161 

0.9930 

0.0232 

0.9966 

-0.0137 

0.9847 

0.0294 

20 

0.0150 

0.9985 

-0.0165 

0.9938 

0.0154 

0.9984 

-0.0161 

0.9898 

0.0141 

25 

2 

0.0870 

0.9592 

-0.0061 

0.9336 

0.1117 

0.9226 

0.0325 

0.8535 

0.1167 

5 

0.0461 

0.9865 

-0.0187 

0.9588 

0.0497 

0.9837 

-0.0109 

0.9413 

0.0494 

10 

0.0329 

0.9924 

-0.0201 

0.9616 

0.0338 

0.9919 

-0.0180 

0.9538 

0.0273 

20 

0.0274 

0.9943 

-0.0198 

0.9610 

0.0273 

0.9943 

-0.0198 

0.9574 

0.0120 

50 

2 

0.0983 

0.9406 

0.0127 

0.8716 

0.1135 

0.9109 

0.0441 

0.8081 

0.1048 

5 

0.0598 

0.9750 

0.0001 

0.8959 

0.0598 

0.9742 

0.0049 

0.8836 

0.0427 

10 

0.0482 

0.9825 

-0.0001 

0.8962 

0.0477 

0.9826 

0.0013 

0.8904 

0.0235 

20- 

0.0411 

0.9868 

0.0017 

0.8927 

0.0409 

0.9868 

0.0027 

0.8893 

0.0128 

398 


For  map  3d,  consisting  of  large  forest  patches  (diameter  100  original  pixels 
each)  and  having  a  total  forest  cover  70%  this  approximation  is  quite  good  even 
for  the  small  block  sizes.  For  degradation  level  10  for  example  the  slope  coefficient 
is  0.9592  for  block  size  2  ('true'  coefficient:  0.9978)  and  0.9976  for  block  size  20, 
being  almost  equal  to  the  'true'  1.0044.  For  Map  la,  consisting  of  very  small 
patches  with  a  total  cover  of  only  about  10%,  there  is  a  considerable  difference 
betweeen  the  slope  coefficient  for  block  size  2  (0.4413)  and  the  true'  value  of 
0.7028,  while  for  block  size  20  the  difference  between  misregistered'  slope 
(0.7416)  and  true'  slope  (0.761 1)  is  small. 

The  differences  in  slope,  however,  do  not  lead  to  an  error  of  estimation  of 
overall  "fcp.  This  is  illustrated  in  Table  3.  The  differences  are  leveled  out  by  the 
adverse  differences  in  intercept:  Lower  slope  coefficients  go  together  with  higher 
intercepts.  In  Map  la  the  differences  in  intercept  are  relatively  small.  But  one  has 
to  consider  that  the  vast  majority  of  observations  in  this  map  is  in  the  class  fcp=0, 
as  total  forest  cover  is  only  10%.  So  small  differences  in  intercept  have  a  large 
effect. 

When  the  calibration  regressions  based  upon  misregistered  blocks  are  used  for 
map  production  (calibrating  the  fcp  of  each  coarse  resolution  pixel  block)  a  clear 
error  is  introduced.  For  larger  fcp  values  the  true  value  will  be  underestimated,  for 
very  small  values  overestimated.  Negative  intercepts  as  they  occur  in  Table  4  are 
normally  not  acceptable  in  a  calibration  regression:  It  would  mean  that  a  forest- 
free  coarse  resolution  pixel  block  is  calibrated  from  an  fcp  of  0.0  to  a  negative  fcp. 
In  our  simulation  study,  however,  the  calibration  regression  was  intentionally  not 
forced  to  be  non-negative. 

The  root  mean  square  error  (rmse)  is  in  most  cases  higher  for  Regression  2 
what  would  be  expected  as  misregistration  introduces  more  variability.  The  rmse, 
however,  is  not  an  adequate  measure  here  as  calculations  were  made  using  per 
block  the  mean  of  the  fcp  of  100  simulations.  Thus,  if  one  is  interested  in  the 
variability  for  one  single  set  of  misregistered  blocks,  one  would  have  to  take 
additionally  into  account  the  variability  within  the  individual  blocks. 

7.  CONCLUSION 

Three  major  factors  can  be  identified  which  determine  the  magnitude  of  the 
misregistration  effect  as  to  this  simulation  study: 

-  Magnitude  of  misregistration.  The  bigger  the  dislocation  distance  the  bigger 
the  difference  in  fcp  values  registered.  This  obvious  relation,  however,  must  also 
be  seen  in  interacting  context  with  the  two  other  factors. 

-  Size  of  pixel  blocks  to  be  registered.  The  bigger  the  block  size  the  less  the 
effect  of  misregistration  For  small  pixel  blocks  the  effect  can  be  considerable.  This 
relation  interacts  very  much  with  the  third  important  factor,  the 

-  spatial  structure  and  fcp  of  the  region  of  interest.  For  very  high  and  very 
low  fcp  the  misregistration  effect  is  smaller  than  for  intermediate  fcp  's.  If  the  forest 
patches  are  large  in  comparison  to  the  size  of  the  pixel  blocks  to  be  coregistered, 
the  effect  of  misregistration  on  fcp  estimation  is  smaller. 

The  bias  of  the  total  forest  cover  estimate  that  was  suspected  on  the  basis  of 
theoretical  considerations  could  not  be  confirmed  to  be  significant  in  this 


399 


simulation  study.  However,  the  effect  of  geometric  misregistration  can 
considerably  add  to  the  error  of  the  estimation  of  total  fcp. 

The  effect  on  map  accuracy  appears  more  important.  Though  the  overall  forest 
cover  estimate  was  found  to  be  almost  not  affected  at  all,  there  was  a  considerable 
effect  on  the  coefficients  of  the  calibration  regression.  This  deviation  of  the 
'misregistered'  calibration  regression  from  the  true  one  leads  to  leads  in  the  cases 
investigated  here  to  an  underestimation  of  high  fcp  values  and  to  an  overestimation 
of  low  ones. 

It  may  be  concluded  that  misregistration  is  one  source  of  error  to  be  considered 
when  simultaneously  analysing  images  of  different  resolution  under  a  double 
sampling  approach. 

ACKNOWLEDGEMENT:  The  authors  thank  Mr.  Olaf  Kugler,  Mr.  Jurgen  Harm  and 
Mrs.  Manuela  Bacher  for  their  help  during  the  computational  procedures. 

REFERENCES 

Cochran,  W.G.  (1977):  Sampling  Techniques.  3rd  edition.  John  Wiley&Sons. 

Ehrlich,  D.,  J.  E.  Estes  and  A.  Singh  (1994):  Applications  of  NOAA-AVHRR  1km  data 

for  environmental  monitoring.  Int.  Journal  of  Remote  Sensing  15(1):  145-161. 
Iverson,  L.  R.;  Cook,  E.  A.  and  Graham,  R.  L.  (1989),  Estimating  Forest  Cover  over 

Southeastern  United  States  using  TM-calibrated  AVHRR  Data.  In  Proceedings  of 

Global  Natural  Research  Monitoring  and  Assessments:  Preparing  for  the  21st 

Century,  Venice,  Italy,  September,  VOL  3,  pp.  1252-1262. 
Justice,  C.  O.,  Markham,  B.  L.,  Townshend,  J.  R.  G.,  and  Kennard,  R.  L.  (1989),  Spatial 

degradation  of  satellite  data.  Int.  Journal  of  Remote  Sensing  10(9):  1539-1561. 
Kleinn,  C,  B.  Traub,  M.  Dees  (1995):  Calibrating  Forest  Cover  Estimates  from  Coarse 

Resolution  Images  with  Samples  of  High  Resolution  Images  -  A  Model  Study  on 

Possible  Effects  of  Misregistration  -.  Paper  presented  at  IUFRO  World  Congress 

1995,  Tampere,  Finland. 
Malingreau,  J. -P.  (1993):  Introduction  to  the  TREES  Project:  Objectives,  Features, 

Achievements  and  Plans.  TREES  Validation  Workshop,  Joint  Research  Centre, 

Ispra,  Italy,  March  9-10,  1993. 
Nelson,  R.  (1989),  Regression  and  Ratio  Estimators  to  Integrate  AVHRR  and  MSS  Data, 

Remote  Sens.  Environ.  30:201-216. 
Paivinen,  R.,  and  Pitkanen,  J.  (1992),  Calibrating  AVHRR  Data  with  Landsat  TM  Data 

for  Tropical  Closed  Forest  Assessment  in  Ghana,  in  Remote  Sensing  and  Permanent 

Plot  Techniques  for  World  Forest  Monitoring,  IUFRO  S. 4. 02. 05  International 

Workshop  Remote  Sensing  and  World  Forest  Monitoring,  edited  by  Lund,  H.  G., 

Paivinen,  R.,  and  Thammincha,  S.,  Pattaya,  Thailand  pp.  95-102. 
Townshend,  J.  R.  G.,  Justice,  C.  O.,  Gurney,  C.  McManus,  J.  (1992),  The  impact  of 

misregistration  on  change  detection.  IEEE  Transactions  on  Geoscience  and  Remote 

Sensing,  Vol.  30,  No.  5:  1054-1060. 
Zhu,  Z.,  and  Evans,  D.  L.(1992),  Mapping  Midsouth  Forest  Distributions  -AVHRR 

satellite  data  and  GIS  help  meet  RPA  mandate.  Journal  of  Forestry  Vol.  90,  No.  12, 

December  1992,  pp.27-30. 


400 


Linear  Mixture  Modeling  with 
Autocorrelated  Errors 

Jayantha  Ediriwickrema1,  Siamak  Khorram2,  Marcia 
Gumpertz3  and  John  Brockhaus4 

Abstract. — The  linear  mixture  model  assumes  that  model  errors  are 
spatially  uncorrelated.  Spatial  continuities  exist  in  most  geographical 
data.  The  data  values  close  to  each  other  are  likely  to  have  similar 
spectral  characteristics.  Especially  in  remotely  sensed  data,  spatial 
autocorrelations  are  present  among  pixels.  Residuals  of  the  Advance 
Very  High  Resolution  Radiometer  data  are  examined  for  any 
deviation  from  the  random  assumption.  The  correlation  among  the 
residuals  of  the  neighboring  pixels  and  Moran's  I  statistic 
demonstrate  the  significance  of  non-random  distribution  of  residuals. 
This  study  develops  an  autocorrelated  error  model  to  calibrate  the 
linear  mixture  model.  Seven  different  autocorrelation  patterns  are 
considered,  and  for  each  pattern,  the  linear  mixture  model  is 
calibrated  and  land  use/land  cover  class  fractions  are  estimated  from 
the  Advance  Very  High  Resolution  Radiometer  data. 


INTRODUCTION 

The  effects  of  spatial  autocorrelation  are  not  clearly  understood  in  unmixing 
mixed  pixels  by  the  linear  mixture  model  (LMM).  Autocorrelated  errors  are 
common  in  remotely  sensed  data  (Campbell  and  Kiiveri  1988).  Campbell  and 
Kiiveri  (1988)  described  point  spread  function  of  the  sensor  and  atmospheric 
effects  as  causes  for  the  spatial  autocorrelation  among  neighboring  pixels.  The 
radiance  values  of  the  neighboring  pixels  may  show  similar  spectral  characteristics 
due  to  scattering  and  similar  land  use/land  cover  (LU/LC)  class  patterns.  These 
effects  become  apparent  with  increasing  coarseness  of  the  spatial  resolution.  The 
autocorrelated  structure  of  remotely  sensed  data  is  widely  used  as  additional 
knowledge  in  image  classification.  Belward  (1992)  used  spatial  attributes: 
contextual  information  and  autocorrelated  errors  to  evaluate  the  Advance  Very  High 
Resolution  Radiometer  (AVHRR)  data  in  environmental  monitoring.  Campbell  and 
Kiiveri  (1988)  discussed  the  advantages  of  using  the  spatial  relations  of  the 
neighboring  pixels  in  image  classification,  but  they  did  not  observe  a  significant 
advantage  due  to  the  spatial  autocorrelation  when  the  spectral  class  separations  were 
high.  Furthermore,  no  significant  improvement  was  noticed  in  overall  accuracy 
when  spatial  autocorrelations  were  included  in  the  classification  process. 

In  the  empirical  method  of  unmixing  mixed  pixels  by  LMM,  calibration 
coefficients  are  estimated  by  multivariate  regression  analysis.  The  calibration 
coefficients  are  usually  estimated  assuming  the  model  errors  are  spatially 
uncorrelated  with  constant  variance.  Iverson  et  al.  (1989)  used  multivariate 
regression  analysis  to  develop  an  empirical  relationship  between  the  AVHRR  data 


1  Graduate  Assistant,  Computer  Graphics  Center,  NCSU,  Raleigh,  NC. 

2  Professor  and  Director,  Computer  Graphics  Center,  NCSU,  Raleigh,  NC,  and  Dean  of  International  Space 
University,  Pare  d'Innovation,  Communaute  Urbaine  de  Strasbourg,  Blvd.  Gonthier  a"  Andernach,  67400  llkirch,  France 

3  Associate  Professor,  Department  of  Statistics,  NCSU,  Raleigh,  NC. 

4  Assistant  Professor,  Mapping,  Charting  and  Geodesy,  Department  of  Geography  and  Environmental  Engineering, 
United  States  Military  Academy.  West  Point,  NY. 


401 


and  the  forest  cover.  Pech  et  al.  (1986)  examined  multivariate  calibration  methods 
to  estimate  vegetation  cover.  In  all  these  studies,  the  model  errors  were  assumed  to 
be  spatially  uncorrected  with  constant  variance. 

Multivariate  regression  models  with  autocorrelated  errors  are  common  in 
statistics.  Nevertheless,  the  spatial  autocorrelation  effects  are  usually  assumed 
insignificant  in  unmixing  mixed  AVHRR  pixels.  Autocorrelated  error  models  are 
widely  used  for  spatially  correlated  data  accounting  for  the  interaction  between 
neighboring  locations.  Upton  and  Fingleton  (1985)  described  a  generalized  least 
squares  (GLS)  regression  method  for  spatially  autocorrelated  models.  Basically, 
models  with  normally  distributed  spatially  autoregressive  errors  fall  in  two 
schemes:  simultaneous  and  conditional  autoregressive  schemes.  The  autocorrelation 
parameters  in  this  study  were  estimated  by  a  maximum  likelihood  method  described 
by  Upton  and  Fingleton  (1985)  based  on  a  simultaneous  autoregressive  scheme. 
Seven  spatial  autocorrelation  patterns  were  compared  with  the  model  that  assumed 
errors  were  spatially  uncorrected. 

METHOD 

The  study  area,  North  Carolina  Piedmont  showed  a  wide  variation  of  LU/LC 
patterns.  The  diversity  of  the  LU/LC  class  patterns  ranged  from  developed  urban  to 
various  distributions  of  residential,  agriculture,  and  forests.  The  frequency  of  the 
local  variation  was  high  within  this  area.  A  significant  amount  of  autocorrelation 
was  anticipated  among  the  AVHRR  pixels  especially  due  to  mixed  pixel  effects. 

Data 

Classified  Landsat  Thematic  Mapper  (TM)  data  and  NOAA  AVHRR  data 
corresponding  to  the  North  Carolina  Piedmont  data  set  used  by  Khorram  et  al. 
(1994)  were  used  in  this  study.  Two  sets  of  ASCII  data  files  were  developed  in  this 
process.  One  set  was  developed  from  the  classified  Landsat  TM  data  and  the  other 
set  was  derived  from  the  AVHRR  data.  The  file  created  from  the  Landsat  TM  data 
described  LU/LC  class  fractions  within  each  one  square  km  area  corresponding  to 
the  AVHRR  pixels.  The  ASCII  files  created  from  the  AVHRR  data  contained  digital 
number  values  in  14  spectral  bands  (Ediriwickrema  1995). 

The  LMM  with  Autocorrelated  Errors 

The  LMM  assumed  that  there  were  no  interactions  between  the  six  LU/LC 
classes.  It  also  assumed  that  the  resultant  radiation  of  each  pixel  could  vary  only 
due  to  random  noise.  Correlation  between  AVHRR  pixel  values  and  atmospheric 
errors  were  also  considered  as  insignificant.  Under  these  assumptions,  composite 
radiation  of  mixed  pixels  was  related  to  the  component  LU/LC  class  fractions  and  to 
the  spectral  values  of  the  LU/LC  class  pure  pixels  by  Eq.  1.  A  pure  pixel  was 
defined  as  a  pixel  that  was  totally  covered  by  only  one  defined  LU/LC  class. 

c 

Ajk  =XFjiRi<c+ajk>  [1] 

i=l 

where  Ajk=  pixel  value  of  the  j*  AVHRR  pixel  in  band  k,  =  fraction  of  the  ith 
LU/LC  class  within  the  jth  AVHRR  pixel,  Rilc  =  calibration  coefficients  of  the  i,h 

LU/LC  class  in  band  k,  ajk  =  random  noise  of  the  jth  AVHRR  pixel  in  the  band  k, 

and  c  =  the  number  of  LU/LC  classes.  The  error  term  "a;i"  was  assumed  to  be 


402 


spatially  correlated.  The  Gauss-Markov  condition  in  the  OLS  regression  is  not 
satisfied  when  the  residuals  are  spatially  correlated.  One  model  for  spatially 
correlated  errors  is  the  simultaneous  autoregressive  error  model.  In  this  model  the 
error  for  one  pixel  depends  upon  the  errors  of  neighboring  pixels.  The  error  vector 

a  for  one  band  having  "n"  pixels  is  ot  =  pWa  +  u  (Upton  and  Fingleton  1985), 

which  gives  unxl  =  (Inxn  - pWnxn )anxl  and  p  is  a  constant  --  autocorrelation 
parameter.  W  is  a  proximity  matrix  in  which  each  row  indicates  which  pixels  are 
neighbors  in  the  autocorrelation  scheme,  u  =  error  vector  that  satisfies  the  Gauss- 
Markov  conditions,  and  I  =  identity  matrix.  Since  the  elements  of  "u"  are 
uncorrected,  the  simultaneous  autoregressive  model  can  be  fitted  by  OLS  after  pre 

multiplying  both  sides  by  (I  -  pW). 

(I™  " pWnxn )Anxb  =  (Inxn  - pWnxn )FnxcRcxb  +  unxb  [2] 

In  other  words,  A *nxb  =  F*nxc  Rcxb  +unxb,  where  A *nxb  =  (Inxn  - pWnxn )Anxb 
and  F*nxc  =  (Inxn -pWnxn)Fnxc.  Then  the  estimates  of  the  calibration  coefficients 

( Rcxb)  are  written  as  follows. 

o       _/p*T    p*     \-l  p*T     a  *  [3] 

cxb       V        cxn         nxc'cxc        cxn  nxb 

A  reasonable  p  value  for  each  spectral  band  was  required  to  estimate  Rcxb.  In  the 

simultaneous  autoregression,  p  was  chosen  such  that  it  maximized  the  log 
likelihood  function  "ln(L)"  in  Eq.  4. 

ln(L)  =  --|lnd2  +lnlI-pWI,  [4] 

where  a2  =  uTu / nand  u  =  (Inxn  - pWm )Anxb  - (Inxn  - pWnxn )Anxb (Upton  and 

Fingleton  1985).  In  Eq.  4  the  determinant  II  -  pWI  is  not  easy  to  compute  for  large 
data  sets.  Upton  and  Fingleton  (1985)  describe  formulas  to  avoid  computational 

complications  as  well  as  a  way  to  compute  II  -  pWI  for  large  data  sets.  These 
formulae  are  applicable  for  only  one  type  of  proximity  matrix,  the  Rook's  case 
(Upton  and  Fingleton  1985).  Therefore,  a  20  column  and  20  row  subset  was 
selected  from  the  upper  half  by  considering  available  system  resources  to  reduce 
computational  burden.  This  subset  was  used  to  calibrate  the  LMM.  A  similar  size 
subset  was  also  selected  from  the  lower  half  of  the  image  to  estimate  LU/LC  class 
fractions,  and  thereby  examine  the  effects  of  autocorrelated  errors  in  the  LMM. 

Weightings 

The  proximity  matrix  (W)  is  constructed  according  to  the  autocorrelation  pattern. 
Upton  and  Fingleton  (1985)  described  three  weightings  named  Rook,  Bishop  and 
Queen.  Rook  weighting  considers  only  adjacent  pixels  within  a  row  or  column  to 
be  neighbors.  The  Bishop  weighting  considered  only  the  cells  that  touch  at  the 
corners  to  be  neighbors,  while  the  Queen  model  considers  all  surrounding  cells, 
i.e.,  touching  corners  and  sides  to  be  neighbors.  In  this  study,  seven  different 
weighting  structures  were  considered  (Ediriwickrema  1995). 


403 


Proximity  matrix 

In  the  proximity  matrix,  the  cells  that  account  for  contiguity  are  assigned  one,  and 
the  remaining  cells  are  assigned  zero.  Each  row  in  the  proximity  is  divided  by  its 
total  so  that  row  totals  are  all  equal  to  one.  For  example,  a  lattice  of  3  x  3  pixels 
results  in  a  9  x  9  proximity  matrix.  The  proximity  matrix  for  the  weighting  Rook's 
case  will  be  similar  to  the  one  that  is  shown  in  (Ediriwickrema  1995).  In  that 
proximity  matrix,  the  boundary  pixels  are  weighted  differently  than  the  other  pixels. 
Only  the  available  surrounding  pixels  are  accounted  for;  there  are  many  other 
possibilities  for  weighting  pixels  on  the  edges.  When  all  pixels  are  not  weighted 
with  the  same  number  of  pixels,  the  scaling  results  in  an  asymmetric  "W"  matrix. 

Analysis  of  Spatial  Continuity  in  the  Model  Residuals 

The  spatial  autocorrelation  structure  of  the  model  residuals  are  commonly 
examined  by  semivariogram  plots  of  the  data,  h-scatter  plots  (Isaaks  and  Srivastava 
1989)  of  the  errors,  Moran's  I  statistic  (Upton  and  Fingleton  1985).  The  residuals 
from  the  regression  of  AVHRR  data  on  the  LU/LC  class  fractions  using  ordinary 
least  squares  (OLS)  were  used  to  analyze  the  spatial  continuity  in  the  model. 
Ediriwickrema  (1995)  describes  seven  different  spatial  autocorrelation  patterns.  The 
spatial  autocorrelation  coefficient  of  the  pixel  residuals  from  an  OLS  fit  of  the  LMM 
to  their  neighboring  (as  defined  for  each  weighting)  pixel  residuals  was  calculated 
for  each  spectral  band.  The  significance  of  spatial  autocorrelation  of  the  OLS 
residuals  was  tested  using  Moran's  I  statistic. 

Autocorrelation  parameters 

The  magnitude  and  the  sign  of  the  autocorrelation  parameters  measure  the  scale 
and  the  direction  of  spatial  continuity.  The  proximity  matrices  are  commonly 

designed  with  row  totals  equal  to  one,  which  restrains  p  to  be  between  - 1  and  1 . 

Haining  (1990)  pointed  out  that  p  is  an  eigenvalue  of  W"\  so  p  must  lie  between 

i/^max  and  l/X^,  where  is  the  largest  eigenvalue  of  W  and  X^n  is  the  smallest 
eigenvalue  of  W.  As  described  in  section  "The  LMM  with  Autocorrelated  Errors," 

the  p  value  for  each  band  was  calculated  such  that  it  maximized  "ln(L)".  In  this 

study,  the  log  likelihood  function  "ln(L)"  values  were  evaluated  for  different  p 

values  ranging  from  -1  to  1  with  an  increment  of  0.01.  The  p  value  corresponding 
to  the  maximum  "ln(L)",  was  selected  as  the  autocorrelation  parameter  estimate  of 
the  respective  spectral  band. 

Calibration  of  the  LMM 

Calibration  coefficients  were  estimated  by  Eq.  5  based  on  spectral  bands.  For 
each  spectral  band,  six  coefficients  were  estimated  representing  all  LU/LC  classes. 

R     =CF*T  F*    V  F*T  A*  [5] 

cxl       V       cxn         nxc'cxc        cxn  nxl 

Each  time  when  the  calibration  coefficients  were  estimated  for  a  band,  residual 
errors  ( unxl)  were  also  calculated  from  the  same  data.  Finally,  the  overall  calibration 

coefficients  matrix  (Rcxb)  and  the  residual  error  matrix  (unxb)  were  developed.  The 
variance-covariance    matrix    of    the    residual    errors    was    calculated  by 


404 


bxb 


=  ubxn  *  unxb  /n .  In  order  to  compare  the  autocorrelated  models  to  the  model 
that  assumed  errors  are  spatially  uncorrelated,  calibration  coefficients  were 
separately  estimated  assuming  the  model  errors  were  spatially  uncorrelated. 

Estimation  of  the  LU/LC  Class  Fractions  from  the  AVHRR  Data 

The  North  Carolina  Piedmont  data  set  was  divided  into  two  portions.  From  the 
upper  portion,  the  subset  (Anxb)  was  selected  to  calibrate  the  LMM  model.  Subset 
(E*mxb)  was  selected  from  the  second  portion  to  estimate  the  LU/LC  fractions  (Gmxc) 
and  assess  the  autocorrelated  errors  in  the  LMM.  Bmxb  =GmxcQfvK  +wm.K,  where 


cxb 


mxb  ' 


Bmxb  =  digital  values  of  "m"  AVHRR  pixels  in  "b"  spectral  bands,  Qcxb  =  calibration 
coefficients  of  "c"  LU/LC  classes  in  "b"  spectral  bands,  and  wmxb  =  residual  error 

matrix.  Let  Qcxb  =  Rcxb  and  assuming  the  measurement  errors  associated  with  Rcxb 

are  insignificant,  the  regression  model  can  be  written  as         =  G^R^  +  w^ 

and  can  also  be  re-written  by  taking  transpose  as  Bbxm  =  Rj^G^  +  wbxm. 

We  propose  a  simple  model  incorporating  spatial  correlation  within  and  between 
spectral  bands.  Let  b  =  Vec(BT)  =  [blT  bT  b3T....bl4T]T,  where  bk  is  an  m  x  1 
vector  of  AVHRR  values  for  the  k*  spectral  band.  Let 


1  = 


a 


1.1 


a 


14,1 


1,2 


14.2 


1,14 


14,14 


[6] 


denote  the  covariance  matrix  among  bands  within  a  pixel.  Let  Ck  =  (I  -pkW)'' 
denote  the  square  root  matrix  of  the  spatial  autocorrelation  for  band  k.  Then  let  the 
covariance  matrix  for  all  bands  and  all  pixels  be 


Var(b)  = 


CiCiau 


C14Cj  <J14 , 


c.cTg 


1,2 


C.C^cr, 


C14C2<j14j2 


'1^-14^1.14 


C|4  C14  C714  14 


[7] 


This  covariance  model  says  that  the  covariance  among  two  bands  in  two  different 
pixels  is  just  the  product  of  the  components  of  spatial  variance  for  each  of  the  two 

bands  and  the  covariance  among  bands  "k"  and  "1"  within  a  pixel  ak].  The  estimated 
GLS  estimator  of  the  LU/LC  class  fractions  for  the  j*  pixel  is  then 

[8] 


g,GLs=(RV-1R1)-1RV-'bJ, 

where  Vj  is  a  14  x  14  matrix  of  estimates  of  Var(b)  corresponding  to  the  jl  pixel 

with  the  elements  of  Z  estimated  by  X  *bxb  =  ubxn  *  unxb  /n ,  and  b}  is  a  14  x  1  vector 
of  AVHRR  values  for  the  f  pixel. 

The  LU/LC  class  fractions  were  found  to  be  best  estimated  by  the  constraint  least 
square  regression  method  (Ediriwickrema  1995).  The  same  method  was  used  in 
this  study  to  estimate  the  LU/LC  class  fractions  from  the  AVHRR  data.  The  LU/LC 
class  fractions  were  estimated  separately  considering  each  autocorrelated  pattern, 
and  assuming  errors  were  spatially  uncorrelated  (Ediriwickrema  1995). 


405 


Validation 

To  examine  the  developments  in  the  autocorrelated  error  model  to  the  OLS  model, 
the  residual  sum  of  squares  was  calculated  for  four  sample  areas.  From  the  upper 
left  of  each  quadrant,  a  20  x  20  sample  area  was  selected. 

RESULTS 

Correlation  of  Model  Residuals 

The  correlation  of  the  model  residuals  among  the  neighboring  pixels  for  each 
weighting  is  described  in  Figure  1.  Except  for  the  bands  11  and  12  generally  all 
other  bands  show  significant  correlation  among  the  neighboring  pixels  in  all  seven 
spatial  autocorrelation  patterns.  Horizontal  and  vertical  spatial  autocorrelation 
patterns  resulted  in  high  correlation  related  to  the  diagonal  spatial  patterns. 


0.6  T 

0.5  A 

c 

0.4 1 

o 

1 

0.3  5! 

t 

0.2 !! 

o 
O 

0.1  SI 

0 .. 

-0.1  i 

OBI 
-B8 


□  B2 
-B9 


AB3     xB4    xB5    0B6  +B7 
OBIO  nBll  AB12  xB13  *B14 


* 

□ 

-A- 


O 
X 


S 


8 


* 

ft 

□ 


o 

4 


B 


D 

Weighting 


E 


Figure  1. — Correlation  of  model  residuals  among  neighboring  pixels  in  each 

weighting  and  each  spectral  band. 

Significance  of  Autocorrelated  Errors 


"e" 


Except  for  the  z-statistics  in  bands  1 1  and  1 2  in  correlation  patterns  "b",  "d", 
and  "f  (Ediriwickrema  1995),  all  other  z-statistics  are  greater  than  the  upper  one- 
tail  critical  value  at  the  95%  confidence  level  (1.645)  (see  Figure  2).  The  z-statistic 
values  in  all  other  spectral  bands  in  all  autocorrelation  patterns  are  clearly  greater 
than  1.645. 


406 


u20  0a^b^cDd[legf^g 

w  164  : 


Bl    B2    B3    B4    B5    B6    B7    B8    B9    BIO  Bll  B12  B13  B14 

Band 

Figure  2. — z-statistics  of  each  AVHRR  band  in  each  autocorrelated  pattern. 


Validation  of  the  Autocorrelated  Error  Model 

The  difference  of  the  sum  of  squares  of  the  residuals  between  the  OLS  and  the 
autocorrelated  error  model  are  shown  in  Figure  3.  The  residual  sum  of  squares  was 
reduced  in  the  autocorrelated  error  model  for  most  of  the  bands  in  all  four 
quadrants.  The  spectral  bands  acquired  in  late  June  and  in  mid  July  show 
remarkable  improvement.  Except  for  bands  five  and  six  in  the  upper  right  quadrant, 
all  other  bands  in  all  quadrants  show  an  improvement  with  the  autocorrelated  error 
model.  The  magnitude  of  the  improvement  is  dependent  on  the  spectral  bands  and 
the  distribution  of  the  LU/LC  class  patterns.  The  data  acquired  in  the  infrared  band 
showed  better  improvement  over  the  visible  band. 


25000-g 

20000 4 
15000-= 
10000 -j 
5000  i 
0 

-5000- 


m 


UL-Quad 

J 


LL-Quad 


LR-Quad 


n 


UR-Quad 


Bl     B2     B3     B4     B5     B6     B7     B8    B9    B10  Bll   B12  B13  B14 


Figure  3. — Difference  of  sum  of  squares  of  the  residuals  between  the  OLS  and 

the  autocorrelated  error  model. 

CONCLUSION 

The  AVHRR  data  have  autocorrelated  errors.  Moran's  I  statistics  and  correlation 
coefficients  among  neighboring  pixels  in  different  spatial  patter  clearly 
demonstrated  the  spatial  association  of  the  AVHRR  data.  The  spatial  autocorrelation 
parameters  were  evaluated,  and  were  corrected  separately  for  each  spectral  band. 
The  magnitude  and  the  sign  of  the  spatial  autocorrelation  parameters  were 
dependent  on  many  factors:  size  of  the  pixel,  the  spectral  band  of  the  image, 


407 


direction  of  the  autocorrelation  pattern  and  the  shape  and  size  of  the  region.  Wilson 
(1992)  observed  lower  autocorrelation  with  the  adjacent  diagonal  pixels  than  with 
the  adjacent  vertical  and  horizontal  pixels.  In  this  study  also,  models  using  diagonal 
pixels  for  the  autocorrelation  patterns  showed  the  least  autocorrelation.  The  Rook 
and  Bishop  models  further  confirmed  this  observation.  The  decrease  of  the  sum  of 
squares  of  residuals  in  the  autocorrelated  error  model  demonstrated  the  advantage  in 
accounting  for  the  autocorrelated  errors.  The  magnitude  of  the  improvement 
certainly  depends  on  the  spectral  band,  acquisition  period,  and  the  distribution  of 
the  LU/LC  class  patterns.  This  study  identified  the  advantage  of  analyzing  the 
effects  of  autocorrelated  errors  in  unmixing  mixed  pixels  using  simulated  data  with 
adequate  background  knowledge.  The  background  knowledge  is  important  to 
understanding  and  interpreting  the  effects  of  autocorrelated  errors  in  unmixing 
mixed  pixels. 

REFERENCES 

Belward,  A.  S.  (1992).  Spatial  Attributes  of  AVHRR  Imagery  for  Environmental 

Monitoring.  International  Journal  of  Remote  Sensing.  13(2):  193-208. 
Campbell,  N.A.,  and  H.T.  Kiiveri,  (1988).  Neighbor  Relations  and  Remotely 

Sensed  Data.  Report  No.  Internal  Report:.  Wembley  6014,  Western  Australia: 

Division  of  Mathematics  and  Statistics,  CSIRO,  pp.  23. 
Ediriwickrema,  DJ.  (1995).  Modeling  and  Analysis  of  AVHRR  Data  for  Biogenic 

Emission  Inventory  System  (BEIS).  Ph.  D.  Dissertation.  Raleigh:  North  Carolina 

State  University,  Raleigh,  NC  27695. 
Haining,  R.  (1990).  Spatial  Data  Analysis  in  the  Social  and  Environmental 

Sciences.  First  ed.  New  York:  Cambridge  University  Press. 
Isaaks,  E.H.,    and  R.M.   Srivastava.  (1989).  An  Introduction  to  Applied 

Geostatistics.  New  York:  Oxford  University  Press. 
Iverson,  L.R.,  E.A.  Cook,  and  R.L.   Graham.  (1989).  A  Technique  for 

Extrapolating  and  Validating  Forest  Cover  Across  Large  Regions.  International 

Journal  of  Remote  Sensing.  10(1 1):  1805-1812. 
Khorram,  S.,  L.  Allen,  J.  Aanstoos,  J.  Brockhaus,  A.  Sampson,  H.M.  Cheshire, 

and   D.J.    Ediriwickrema,    (1994).    Vegetation    Canopy    and    Land  Use 

Characterization.  Report  No.  #68D20009:  Final  Report  Submitted  to  U.  S. 

Environmental  Protection  Agency,  Research  Triangle  Park,  NC  277 1 1 .  Raleigh, 

NC  27695-7106:  Computer  Graphics  Center,  North  Carolina  State  University, 

pp.  54. 

Pech,  R.P.,  A.W.  Davis,  R.R.  Lamacraft,  and  R.D.  Graetz.  (1986).  Calibration  of 
LANDSAT  Data  for  Sparsely  Vegetated  Semi-Arid  Rangelands.  International 
Journal  of  Remote  Sensing.  7(12):  1729-1750. 

Upton,  G.J.G.,  and  B.  Fingleton.  (1985).  Spatial  Data  Analysis  by  Example.  Vol. 
1.  Chicheter,  New  York:  John  Wiley  &  Sons,  2  vols. 

Wilson,  J.D.  (1992).  A  Comparison  of  Procedures  for  Classifying  Remotely- 
Sensed  Data  Using  Simulated  Data  Sets  Incorporating  Autocorrelations  Between 
Spectral  Responses.  International  Journal  of  Remote  Sensing.  13(14):2701- 
2725. 


408 


Alternative  Method  to  Validate  the 
Seasonal  Land  Cover  Regions  of  the 
Conterminous  United  States 

Zhiliang  Zhu1,  Donald  O.  Ohlen1,  Raymond  L.  Czaplewski2, 

and  Robert  E.  Burgan3 


Abstract. —  An  accuracy  assessment  method  involving  double  sampling 
and  the  multivariate  composite  estimator  has  been  used  to  validate  the 
prototype  seasonal  land  cover  characteristics  database  of  the  conterminous 
United  Stales.  The  database  consists  of  159  land  cover  classes,  classified 
using  time  series  of  1990  1-km  satellite  data  and  augmented  with 
ancillary  data  including  terrain,  climate,  and  ecological  definitions. 
Reference  data  from  more  than  2,200  1  km2  plots  were  independently 
collected;  data  included  descriptions  of  dominant  land  cover  types. 
Validity  of  the  database  is  presented  in  terms  of  conditional  probabilities 
and  variability  estimates.  This  case  study  demonstrates  the  usefulness  of 
the  method  in  describing  the  reference  composition  of  the  mapped 
categories,  handling  different  classification  schemes  between  the  database 
and  reference,  and  comparing  various  regrouping  complexities.  The 
validation  results  support  using  the  database  as  a  vegetation  baseline  in 
large  scale  environmental  simulation  models. 


INTRODUCTION 

Data  acquired  by  the  National  Oceanic  and  Atmospheric  Administration's 
advanced  very  high  resolution  radiometer  (AVHRR)  represent  one  of  the  most 
important  data  sources  for  global  change  research  (Townshend  1994).  The  daily 
coverage  of  a  large  area  and  1-km  spatial  resolution  (at  nadir)  of  AVHRR  allow 
users  to  effectively  and  efficiently  map  and  study  vast  areas,  at  scales  from 
continental  to  global.  An  example  is  a  recent  prototype  land  cover  characteristics 
database  for  the  conterminous  United  States  (Brown  and  others  1993,  Loveland 
and  others  1993,  1995). 

The  conterminous  U.S.  (CONUS)  land  cover  characteristics  database,  produced 
by  the  U.S.  Geological  Survey's  EROS  Data  Center,  in  cooperation  with  the 
University   of  Nebraska-Lincoln,   defines    159   seasonally   distinct  classes. 


'Senior  Scientist,  Hughes  SIX  Corporation,  LROS  Data  Center,  Sioux  Falls,  SD  57198. 

Mathematical  Statistician,  USDA  Forest  Service.  Rocky  Mountain  Forest  and  Range  F.xpcriment  Station,  Fort 
Collins.  CO  80526. 

^Research  Forester.  USDA  Forest  Service,  Intennounfain  Research  Station,  Missoula,  MT  59807. 


409 


Vegetation  seasonal  properties,  or  phenology,  are  defined  by  a  time  series  of  the 
normalized  difference  vegetation  index  (NDVI),  a  transformation  of  the  first  two 
AVHRR  spectral  bands.  Each  class  in  the  CONUS  database  exhibits  a  unique 
combination  of  seasonal  characteristics  and  landscape  pattern,  such  as  vegetation, 
climate,  and  elevation.  However,  the  concept  and  use  of  polythetic  regionalization 
(Loveland  and  others  1995)  imply  that  the  classes  consist  of  a  mixture  of  land 
cover  types  within  each  class. 

The  CONUS  land  cover  characteristics  database  has  been  widely  used  in  large- 
scale  environmental  simulation  modeling  (Steyaert  and  others  1994).  The  concept 
and  methods  of  the  prototype  database  are  also  being  used  to  produce  similar  land 
cover  characteristics  databases  for  the  major  continents  of  the  world  (Loveland 
and  others  1995).  Because  of  the  current  utility  and  the  future  applications, 
validation  of  the  CONUS  database  is  an  important  issue.  Loveland  and  others 
(1995)  discussed  the  importance  of  validation  and  various  validation  plans, 
including  a  study  reported  by  Merchant  and  others  (1994).  Preliminary  results 
from  that  study  indicated  that  the  database  closely  represented  land  cover  regions 
of  the  conterminous  United  States.  The  objective  of  this  research  is  to  test  an 
alternative  method  for  validating  the  CONUS  database,  as  presented  below. 

ACCURACY  ASSESSMENT  ALTERNATIVES  FOR  CONUS 

Accepted  conventional  methods  for  accuracy  assessment,  as  reviewed  in 
Congalton  (1991),  require  that  reference  data  and  the  remote  sensing  product  in 
question  share  a  common  classification  scheme,  i.e.,  have  the  same  number  of 
corresponding  classes.  Positional  accuracy  for  the  map  product  is  assessed  with 
a  square  error  matrix  that  shows  percent  sample  points  correctly  classified,  where 
the  sample  points  are  geographically  coincident.  The  construction  of  error 
matrices  also  allows  calculation  of  the  kappa  statistic  (Congalton  1991). 

The  error  matrix  and  kappa  statistic  approach  is  routinely  used  to  validate  land 
cover  classifications  with  moderate  to  fine  ground  resolutions.  This  method  has 
not  been  effective  for  coarse  resolution  products,  such  as  those  that  are  based  on 
AVHRR  data,  for  several  reasons.  First,  the  1  km2  pixel  on  the  ground  often 
contains  a  mosaic  of  different  cover  types;  finding  corresponding  reference  classes 
at  this  resolution  and  scale  is  difficult.  Second,  requirements  of  the  error  matrix 
restrict  the  use  of  reference  data  acquired  under  a  classification  scheme  different 
from  the  one  used  for  the  satellite  map.  Third,  image  geometry  is  complex  and 
identification  of  ground  samples  is  difficult  or  speculative.  Additionally,  the 
original  design  philosophy  (Loveland  and  others  1991)  of  the  CONUS  database 
was  a  flexible  database  of  land  characteristics,  classes  were  meant  to  be  translated 
or  combined  in  different  ways  to  meet  the  needs  of  a  wide  range  of  applications 
(Steyaert  and  others  1994).  A  single  reference  data  set  would  be  insufficient  and 
difficult  for  comparing  at  all  levels  of  complexities. 

The  consideration  in  this  study  for  validation  of  the  CONUS  database  was  that 


410 


the  reference  data  and  validation  method  would  be  designed  to  describe  reference 
land  cover  compositions  of  the  mapped  land  cover  classes,  thereby  allowing 
evaluation  of  the  database  at  different  levels  and  providing  measures  of  map  data 
quality  to  users.  The  method  needs  to  be  flexible  and  practical  to  permit  use  of 
a  common  reference  for  different  classification  schemes  and  to  incorporate 
multiphase,  multistage  designs  for  improved  efficiency. 

The  alternative  method  used  in  this  study  to  validate  the  CONUS  land  cover 
characterization  database  is  that  of  the  multivariate  composite  estimator 
(Czaplewski  1992).  The  formulation  for  this  approach  is  given  in  detail  in 
Czaplewski  (1992,  1994,  1995);  a  brief  overview  follows. 

Assume  that,  as  a  special  case  of  multiphase  sampling,  a  set  of  satellite- 
classified  (map)  pixels  and  reference  points  is  obtained.  The  proportions  of  each 
category  in  the  sampled  map  and  reference  data  are  expressed  with  their  respective 
state  vectors  and  estimated  covariance  matrices.  The  multivariate  composite 
estimator  is  a  linear  transformation  matrix  of  the  independent  map  and  reference 
state  vectors,  with  a  covariance  matrix  calculated  as  the  linear  function  of  map  and 
reference  covariance  matrices. 

From  the  multivariate  composite  estimator,  two  very  useful  vectors  are  derived. 
The  first  is  a  vector  of  joint  probabilities  (i.e.,  a  vectorized  contengency  table) 
showing  unbiased  estimates  of  proportions  of  the  map  and  reference  categories 
occurring  together  (Czaplewski  1992).  The  second  is  a  nonlinear  transformation 
vector  of  estimated  conditional  probabilities,  showing  estimated  true  ground 
proportions  given  each  map  category.  The  approximate  variance  estimator  for  the 
conditional  probabilities  is  given  in  Czaplewski  (1994). 

The  important  features  of  the  above  work  are  that  it  provides  a  means  of 
statistically  describing  reference  composition  of  coarse  resolution  map  categories, 
and  the  descriptions  can  be  done  at  various  levels  of  detail  with  the  same 
reference  data  set.  The  approach  accommodates  the  situation  where  different 
classification  schemes  are  used  by  the  remote  sensing  product  and  the  reference 
data  (Czaplewski  1992),  and  it  allows  for  more  complex  sampling  designs 
(Czaplewski  1994)  beyond  the  simple  random  sample  of  point  plots  required  by 
the  conventional  method  (Congalton  1991). 

The  U.S.  Forest  Service  has  produced  a  software  program  based  on  this 
approach  (Williams  and  Beach  1995).  The  program  produces  the  joint  probability 
vector,  the  conditional  probability  vector,  and  the  confidence  intervals  calculated 
using  the  estimated  variance  and  the  beta-binomial  distribution  (Czaplewski  1995). 

SAMPLING  PROTOCOL  AND  REFERENCE  PLOTS 

Congalton  (1991)  states  that  the  sample  size  for  reference  data  used  to  assess 
accuracies  of  remote  sensing  products  is  often  dictated  by  practical  considerations. 
In  the  case  of  the  CONUS  database,  the  study  area  (conterminous  United  States) 
is  very  large,  the  number  of  classes  (159)  is  high,  and  the  AVHRR  spatial 


411 


resolution  makes  reference  data  expensive.  Because  of  these  constraints,  the 
sampling  design  described  below  was  a  process  evolving  between  the  original 
design  (Burgan  and  others  1993)  and  practical  compromises  at  later  stages. 

A  two-stage,  stratified  sampling  design  (Burgan  and  others  1993)  was  used  to 
select  698  primary  sampling  units  (PSU)  and  3,500  secondary  sampling  units 
(SSU).  The  PSU's  were  USGS  7.5  minute  quadrangle  maps  and  the  SSU's  were 
1-km2  field  plots  corresponding  to  1-km2  pixels  in  CONUS  database.  A  decision 
was  made  not  to  sample  classes  with  fewer  than  5,000  pixels,  as  well  as  urban 
areas  and  water.  Field  work  was  coordinated  by  the  U.S.  Forest  Service  (Burgan 
and  others  1993)  and  carried  out  by  personnel  from  several  Federal  and  State 
agencies.  Because  of  various  difficulties,  the  actual  number  of  SSUs  measured 
as  reference  materials  was  2,284.  This  accounted  for  a  0.03  percent  sample 
intensity  for  7.6  million  pixels,  from  129  classes,  in  the  database.  According  to 
the  procedure  for  reference  data,  each  1-km2  SSU  selected  in  a  PSU  was 
transferred  to  aerial  photographs,  interpreted,  and  then  field-checked  using  a  field 
form  designed  for  the  project.  More  than  100  reference  variables  were  collected 
or  calculated,  including  plot  coordinates,  land  cover  composition,  and  primary  and 
secondary  vegetation  cover  types  or  species. 

Among  these  variables,  the  most  pertinent  to  this  validation  were  those  of  17 
dominant  cover  types  (listed  as  the  reference  classification  at  the  top  of  the  figures 
1-4).  Within  each  of  the  1  km2  plots,  percent  values  of  land  cover  by  the  17 
cover  types  were  tallied  and  summed  to  100  percent.  For  example,  the  reference 
data  for  plot  number  3  was  20-percent  conifer  with  crown  closure  less  than  30 
percent,  20-percent  conifer  with  crown  closure  greater  than  60  percent,  50-percent 
mixed  forest  with  crown  closure  less  than  30  percent,  and  10-percent  shrub.  The 
CONUS  map  classification  for  plot  3  was  western  U.S.  conifer  forests  of 
evergreen  needleleaf  species.  The  17  variables  and  the  2,284  reference  points 
were  used  for  comparisons  to  the  CONUS  database  classes. 

RESULTS  AND  DISCUSSION 

The  reference  data  set  was  used  to  assess  several  categorical  simplifications 
(groupings)  of  the  CONUS  database  to  validate  the  map  for  applications. 
Conditional  probability  vectors,  standard  errors,  and  confidence  intervals  at  the  80 
percent  confidence  level  were  obtained  for  the  following  simplications:  6  land 
cover  classes  (figure  1),  12  land  cover  classes  (figure  2),  a  modified  Anderson 
level  II  classification  (Anderson  and  others  1976)  (figure  3),  and  a 
Biosphere- Atmosphere  Transfer  Scheme  (BATS)  (Dickinson  and  others  1986) 
translation  (figure  4).  Actual  data  for  the  four  simplifications  as  well  as  the  full 
129  CONUS  classes  are  also  available  from  the  authors. 

In  the  figures,  for  each  class  category,  conditional  probabilities  of  the  17 
reference  dominant  cover  types  are  shown  with  80  percent  confidence  intervals. 
For  example,  in  figure  1,  if  the  CONUS  class  is  deciduous  forests,  the  actual 


412 


Reference  classification 


Foras!  lands 
Conitoroul  tAxmd         5  3 


8^ 


50% 


Agricultural 
land 


33.1%  wzzzmiB 


0% 


50% 

Grassland 

12.2%  W7M 


o% 


V 


1 


50% 

Shrubland 

18.8%  W/MM 


*  * 


o%«*t 


50% 

Coniferous 
Forest 

12.2% 


0% 


50% 

Deciduous 
Forest 

11.3% 


1         „  • 


50% 


Mixed 
Forest 

12.4% 


Reference  classification 


For »st  lands 
Contorous  Miiatf 
D*cxJuous 


Forssl  lands 
Conrfarous  rAx«d 

D»csfluou»  ^ 


1 

50% 

I  |!  si;  !  ;|  ;| 

Cropland 

15  .1%  w/////( 

•^mi  &  s;  ©  ss  s; 

i  | 

0% 

50% 


Mixed 

1 2.6% 


I  ,|  |lf  i 

o%§;  :  ■    .    .  • 


50% 

Cropland/  . 
Grassland 

5.7% 


50% 


o% 


Cropland/ 
Woodland 

13.0%  mm 


I 


50% 

Grassland 

9.1%  VmTMMA 


0% 


50% 


I1  1  .* 


0% 


50% 

Savanna 

2.9%  m 


o% 


50% 

Shrubland/ 
Grassland 

12.3% 


.  Chaparral 

j    I       :     0.4%  \ 

in 


o% 


50% 

Desert/ 
Shrubland 

7.0% 


o% 


* 

0% 

■JJJ 

50% 

Deciduous 
Forest 

9.5%  W/Wk 


Coniferous 
Forest 

10.0%  Wffl&fi 


o% 

50% 

Woodland 

2.5%  m 


0% 


o%« 


i  •  • ,  • 


III 


IB 


Figure  I.  --  Probability  of 
reference  classification  given 
collapsed  CONUS  classfication 
system  with  6  categories. 
Horizontal  bars  show  prevalance 
of  CONUS  mapped  categories. 


Figure  2.  —  Probability  of  reference  classification  given 
collapsed  CONUS  classfication  system  with  12  categories. 
Vertical  lines  display  80%  confidence  interval  for  the  sampling 
error  from  the  698  sample  quadrangles.  Horizontal  bars  show 
prevalance  of  CONUS  mapped  categories. 


413 


Forest  lands 
xis  K 
Deciduous 


  -  -j*.  ?  1  i    |  :  | 

s:S2S8?s:s 2^5 111*?*  si 


Reference  classification 

j  I  fill  s 
Palis 


F  of  est  lands 
Coniferous  Mixed 
Deciduous 


)  XD  w- 

}  O  6  .  _  _  _ 


'11 


IP! 


Forest  lands      Vt.  ™. 
Coniferous  Mixed :-x-  g 


3n  JTg  2 


SS'8 


00-0  o  o 
006 


I  gjB-g 


»:■>:■  -fi 
?  □ 

?r  z  M-.  2 


50% 


Irrigated 

Cropland 

and 

Pasture 


0.6% 


0% 


50% 

Grassland 

Cropland 

Mosaic 


Mil 


siiJ-i 


5.5% 


0% 


Ml 


50% 

Subalpine 
Forest 

0.1%  8 

0% 


Dryland 

Cropland 

and 

Pasture 


13.4% 


0% 
50% 

Herbaceous 
Coastal 
Wetlands 

0.1%  a 

0% 

50% 

Mixed 
Dryland/ 
Irrigated 
Cropland 

0.6%   0  Oo/o 

50% 

Grassland 

8.8% 


Siii: 


50°/< 

Barren  or 

Sparsely 

Vegetated 

2.5%  m 

0%, 
50% 

Forested 
Coastal 
Wetlands 

0.1%  Q 

0% 

50% 

Desert 
Shrubland 

e.8%  m 


li 


50% 


Alpine 
Tundra 

0.1%  0 


0% 

50% 

Western 
Coniferous . 
Forest 

7.8% 


Mixed 
Shrubland/ 
Grassland 

12.0% 


0% 

50% 

Western 
Woodlands 

2.5%  m 


o^i:rii>'a.M:;  ate: 


Western 
Deciduous 
Forest 

0.1%  8 

0% 

50% 

Northern 
Deciduous . 
Forest 

3.6%  WA 


0% 

50% 

South- 
eastern 
Deciduous 
Forest 

5.5%  V//////A 


50% 

Chaparral 

\     0.4%  @ 


0% 

50% 

Woodland/ 

Cropland 

Mosaic 

12.7% 


♦  ON 


i.l 


50% 


Western 

Mixed 

Forest 

1.0%  0 


0% 

50% 


50% 


Savanna 

2.9%  vm 


0% 


0% 
50% 

South- 
eastern 
Coniferous. 
Forest 

$*&    2.0%  YA  0% 


i 

1 

1  It; 

Northern 

Mixed 

Forest 

3.3%  WM 


11 


0% 

50% 


1 

•  1  :i 

M 

it  j 

IJ  M 

LI 

III 


South- 
eastern 
,  Mixed 
.  r  ,|. ,  Forest 


liii  » 


Figure  3.  —  Probability  of  reference  classification  given  collapsed  CONUS  classfication  system 
that  approximates  a  modified  Anderson  Level  II  system.  Vertical  lines  display  80%  confidence 
interval  for  the  sampling  error  from  the  698  sample  quadrangles.  Horizontal  bars  show  prevalance 
of  mapped  Anderson  categories. 


414 


For  sal  lands 
Conilarous  Mu«d 
D*aduous 


II  i 
s  s  2 


28  S  2  8  S  °  8  SjS  <S  5      I  £  i 


Reference  classification 

Forest  lands  o 
Conilorous  Musd  *  £  g 

D»oduous  f  *     3  f 


Foresl  lands 
Coralwous  Miad 
Dtaduous 


I  s 


1! 


50% 

Cropland/ 
mixed 
farming 

13.6% 


50% 

Short  grass 
cropland 

4.6%  ffl 

o% 

50%, 

Short  grass 
grassland 

7.8%  WZA 

o% 


50% 

Tall  grass 
grassland 

0.6%  g 

o% 

50% 

Evergreen 
needleleaf 
forest 

9.7%  1 

50% 

Evergreen 
shrubland 


I 


50%; 


i 


Irrigated 
cropland 

1.2%  g 


i 


507c 

Evergreen 

needleleaf/ 

broadleaf 

0  1%  I 

o% 


♦  r 

U  Mlrtf 


50% 

Tall  grass/ 
cropland 

1.2%  0 


I 


I 


Bog  or 
marsh 

0.4%  a 


o%| 


■  •  •  •  • 


50% 

Semidesert 

3.0%  0 

o% 


1 1 1 1 1 1 ; 


50% 


0% 
50% 


il  t 


.1 


Mixed 
woodland 

9.2% 


50% 


Desert 

2.5%  0 


0% 
50% 


Tundra 

0.1%  B 


o% 


50% 


0% 


km 


»::.  :f::  "a  ft  i:  y  •  « 


Deciduous 
broadleaf 
forest 

9.2% 


sua 


15.7%, 


0%«  « 


PI 
i*i — 3iL 


50% 

Cropland/ 
evergreen 
needleleaf 
forest 

1.3%  2  o% 
50%. 

Evergreen 
needleleaf/ . 
short  grass 

2  6%  0 

o% 


n  1 

50%, 

Deciduous  j 
shrubland 

0.7%.  S 

o% 

50% 

Cropland/ 
deciduous 
broadleaf 
forest 

7  .1%  MA 

50% 

Cropland/ 

mixed 

woodland 

6.8%  Wh 

o% 


s 


AS 


50% 

Irrigated 
crop/ 

deciduous 
broadleaf  L 

0.3%.  \  0%j 


50% 

Short 
grassland/ 
mixed 

|        i  woodland 

1  ill  26%  0  0% 


ill 


jMILMi 


Figure  4.  --  Probability  of  reference  classification  given  collapsed  CONUS  classification  system 
that  approximates  the  Biosphere- Atmosphere  Transfer  Scheme  (BATS).  Vertical  lines  display 
80%  confidence  interval  for  the  sampling  error  from  the  698  sample  quadrangles.  Horizontal 
bars  show  prevalance  of  mapped  BATS  categories. 


415 


cover  types,  according  to  reference  field  plots,  are  approximately  40-percent 
deciduous  forests  with  crown  closure  greater  than  60  percent,  14-percent 
deciduous  forests  with  crown  closure  between  30  and  60  percent,  11 -percent 
agricultural  lands,  10-percent  grassland,  and  the  remaining  25-percent  other  cover 
types.  In  other  words,  the  collapsed  CONUS  class,  despite  its  oversimplified  label 
as  deciduous  forest,  is  a  region  of  pixels  with  distinct  seasonal  properties  but  not 
necessarily  the  conventional,  mutually  exclusive  land  cover  properties  (Loveland 
and  others  1995).  This  seasonal  region  is  expressed  as  a  function  of  the  17- 
element  probability  vector;  values  of  the  17  elements  indicate  the  composition  of 
the  region.  In  the  example  above,  the  "deciduous"  seasonal  region  is  composed 
of  a  majority  of  deciduous  forests  (55  percent),  plus  agricultural  and  grass  land 
cover. 

The  grassland  class  is  similarly  expressed  in  figure  1  as  a  seasonal  region 
consisting  of  not  only  "pure"  grass  land  cover  (30  percent),  but  also  a  mixture  of 
grass  and  shrub,  crop,  and  forests,  as  well  as  nonvegetated  land.  Interpreting  the 
results  requires  an  understanding  of  the  interspersion  of  crop  and  forest  land  cover 
in  some  areas  that  may  resemble  grassland  in  terms  of  their  spectral  and  temporal 
characteristics.  Misclassification  of  nonvegetated  land  caused  by  factors  such  as 
AVHRR  data  quality  (low  sun  angle,  viewing  geometry,  atmospheric  effects)  may 
also  partially  contribute  to  the  results. 

The  standard  error  and  the  confidence  interval  are  measures  of  sampling  error 
for  individual  estimated  probabilities.  These  variability  terms  can  be  used  in 
large-scale  environmental  simulation  models  for  calculating  error  propagation.  It 
should  be  noted  that  when  sample  size  is  small  for  an  element  of  the  probability 
vector,  the  variance  estimators  may  not  be  reliable.  In  this  case,  the  variance 
measures  should  be  noted  but  not  used. 

Examination  of  figures  1  through  4  shows  the  flexibility  of  the  alternative 
validation  method  in  using  common  reference  set  at  different  levels  of  CONUS 
simplification.  Note  that  when  CONUS  classes  are  recoded  or  translated,  the 
constituent  classes  are  changed.  For  example,  grassland  category  in  the  6  class 
scenario  (figure  1)  includes  savanna  class,  but  grassland  is  separate  from  savanna 
in  the  12  class  scenario  (figure  2),  and  it  is  split  into  short  and  tall  grass 
subcategories  in  a  BATS  translation  (figure  4).  Compositions  of  deciduous  forests 
and  other  categories  are  similarly  changed.  However,  as  shown  in  figures  1 
through  4,  their  validity  in  terms  of  probabilities  of  ground  reference  and  standard 
errors  generally  holds  throughout  the  simplifications  of  classification  detail. 
Deciduous  classes  maintain  similar  constituencies  except  the  northern  deciduous 
forest  class  within  the  Anderson  translation,  where  the  seasonal  region  contains 
a  larger  share  of  agricultural  lands. 

SUMMARY 

Land  cover  classifications  and  databases  using  coarse  resolution  AVHRR  data 


416 


represent  important  inputs  into  environmental  simulation  models  for  global  change 
research.  Validation  of  AVHRR  based  products  is  critical  but  is  often  not 
effectively  fulfilled  because  of  the  mismatch  between  conventional  methods  for 
accuracy  assessment  and  characteristics  of  AVHRR  data. 

An  alternative,  and  complementary,  approach  involves  using  conditional 
probability  and  variability  terms  to  describe  the  nature  and  validity  of  the  CONUS 
land  cover  characterization  database.  It  is  important  to  realize  that  all  land  cover 
classifications  contain  errors.  As  stated  before,  CONUS  classes  are  different  from 
conventional,  monothetic  spectral  classifications  in  that  the  CONUS  classes  are 
generally  dominated  by  one  land  cover  type,  but  at  the  same  time  are  comprised 
of  other  types  of  land  cover  that  share  similar  seasonal  and  spectral  characteristics. 
Conditional  probabilities  and  error  terms  from  this  study  show  that  the  CONUS 
classes  accurately  described  such  landscape  patterns  and  are  generally  consistent 
across  various  levels  of  classifications  and  translations. 

Additional  research  on  the  validation  of  large  area  land  cover  mapping  should 
look  at  the  correlations  of  spatial  components.  Evaluating  spatial  errors  found 
between  the  probabilities  of  dominant  reference  cover  types  and  the  various 
CONUS  land  cover  mappings  should  provide  an  understanding  into  any 
regionalization  of  land  cover  seasonal  characteristics. 

The  validation  method  used  in  this  study  is  flexible  enough  to  handle  different 
classification  schemes  for  map  and  reference  categories  and  complex  sampling 
designs,  and  the  same  reference  data  set  can  be  used  to  evaluate  map  categories 
translated  at  various  levels  of  complexity.  Therefore  this  approach  of  validation 
and  accuracy  improvement  for  area  estimates  of  different  land  cover  categories  is 
useful  in  land  cover  mapping  projects  of  the  global  scale. 

ACKNOWLEDGMENTS 

The  research  described  in  this  article  has  been  supported  by  the  U.S. 
Environmental  Protection  Agency  (EPA)  through  Interagency  Agreement  IAG 
DW 14936073  to  the  U.S.  Geological  Survey.  However,  it  has  not  been  subjected 
to  EPA  review  and  therefore  does  not  necessarily  reflect  the  views  of  the  agency. 
No.  official  endorsement  should  be  inferred.  The  Hughes  STX  Corporation 
performed  this  work  under  U.S.  Geological  Survey  contract  1434-92-C-40004. 

REEERENCES 

Anderson,  J.R.,  Hardy,  E.E.,  Roach,  J.T.,  and  Witmer,  R.E.  1976.  A  land  use  and 
land  cover  classification  system  for  use  with  remote  sensor  data.  U.S. 
Geological  Survey  Prof.  Paper  964,  28  p. 

Brown,  J.F.,  Loveland,  T.R.,  Merchant,  J.W.,  Reed,  B.C.,  and  Ohlen,  D.O.  1993. 
Using  multisource  data  in  global  land  characterization:  concepts,  requirements, 
and  methods.  Photogrammetric  Engineering  and  Remote  Sensing,  59:977-987. 


417 


Burgan,  R.E.,  Hardy,  C,  Ohlen,  D.O.,  Loveland,  T.R.,  Brown,  J.F.,  and  Reed, 

B.C.  1993.  Land  characterization  map  evaluation,  study  plan  no.  4401-52,  U.S. 

Forest  Service,  24  p. 
Congalton,  R.G.  1991.  A  review  of  assessing  the  accuracy  of  classifications  of 

remotely  sensed  data.  Remote  Sensing  of  Environment,  37:  35-46. 
Czaplewski,  R.L.  1992.  Accuracy  assessment  of  remotely  sensed  classifications 

with   multi-phase   sampling   and  the   multivariate  composite  estimator. 

Proceedings  of  16th  International  Biometrics  Conference,  Hamilton,  New 

Zealand,  December  7-11,  1992.  2:22. 
Czaplewski,  R.L.  1994.  Variance  approximations  for  assessments  of  classification 

accuracy.  Research  Paper  RM-316.  U.S.  Department  of  Agriculture,  Forest 

Service,  Rocky  Mountain  Forest  and  Range  Experiment  Station,  Fort  Collins, 

CO.  29  p. 

Czaplewski,  R.L.  1995.  Assessment  of  classification  accuracy  and  extent  estimates 
for  a  land  cover  map  with  double  sampling.  Submitted  to  Forest  Science. 

Dickinson,  R.E.,  Henderson-Sellers,  A.,  Kennedy,  P.J.,  and  Wilson,  M.F.  1986. 
Biosphere-atmosphere  transfer  scheme  (BATS)  for  the  NCAR  community 
climate  model.  NCAR  Technical  Note  NCAR/TN-275+STR,  Boulder,  CO. 

Loveland,  T.R.,  Merchant,  J.W.,  Ohlen,  D.O.,  and  Brown,  J.F.  1991.  Development 
of  a  land  cover  characteristics  database  for  the  conterminous  U.S. 
Photo grammetric  Engineering  and  Remote  Sensing,  57:1453-1463. 

Loveland,  T.R.,  Ohlen,  D.O.,  Brown,  J.F.,  Reed,  B.C.,  Merchant,  J.W.,  and 
Steyaert,  L.T.  1993.  Prototype  1990  conterminous  United  States  land  cover 
characteristics  data  set  CD-ROM,  USGS  CD-ROM  Set  9307. 

Loveland,  T.R.,  Merchant,  J.W.,  Reed,  B.C.,  Brown,  J.F.,  and  Ohlen,  D.O.  1995. 
Seasonal  land  cover  regions  of  the  United  States.  Annals  of  the  Association  of 
American  Geographers,  85(2):339-355. 

Merchant,  J.W.,  Yang,  L.,  and  Yang,  W.  1994.  Validation  of  continental  scale 
land  cover  data  bases  developed  from  A  VHRR  data.  Proceedings  of  the  Pecora 
12  Symposium  on  Land  Infomiation  from  Space-Based  Systems,  American 
Society  of  Photogrammetry  and  Remote  Sensing,  Bethesda,  MD,  pp.  63-72. 

Steyaert,  L.T.,  Loveland,  T.R.,  Brown,  J.F.,  and  Reed,  B.C.  1994.  Integration  of 
environmental  simulation  models  with  satellite  remote  sensing  and  geographic 
information  systems  technologies:  case  studies.  Proceedings  of  Pecora  12 
Symposium  on  Land  Information  from  Space-Based  Systems,  American  Society 
of  Photogrammetry  and  Remote  Sensing,  Bethesda,  MD,  pp  407-417. 

Townshend,  J.R.G.  1994.  Global  data  sets  for  land  applications  from  the 
Advanced  Very  High  Resolution  Radiometer:  an  introduction.  International 
Journal  of  Remote  Sensing,  15(17):33 19-3332. 

Williams,  M.T.,  and  Beach,  D.J.C.  1995.  ACAS  0.4:  accuracy  assessment  system 
program  manual.  U.S.  Department  of  Agriculture,  Forest  Service,  Rocky 
Mountain  Forest  and  Range  Experiment  Station,  Fort  Collins,  CO.  33  p.  + 
source  code. 


418 


Investigation  of  Possible 
Contributions  NDVFs  Have  to 
Misclassification  in  AVHRR  Data 

Analysis1 

David  L.  Evans2  and  Raymond  L.  Czaplewski3 

Abstract. — Numerous  subcontinental -scale  projects  have  placed 
significant  emphasis  on  the  use  of  Normalized  Difference  Vegetation 
Indices  (NDVTs)  derived  from  Advanced  Very  High  Resolution 
Radiometer  (AVHRR)  satellite  data  for  vegetation  type  recognition.  In 
multi-season  AVHRR  data,  overlap  of  NDVI  ranges  for  vegetation 
classes  may  degrade  overall  classification  performance.  The  preliminary 
work  of  this  project  is  presented  to  illustrate  potential  differences  in 
vegetative  cover  discrimination  capabilities  between  original  visible  and 
near-infrared  (NIR)  channels  of  AVHRR  data  and  NDVTs  derived  from 
these  channels.  Initial  tests  on  AVHRR  data  have  been  performed  for 
regions  of  known  vegetative  cover  composition  in  Northern  Mexico.  The 
results  are  presented  here  to  stimulate  debate  on  the  effectiveness  of  use  of 
NDVTs  for  vegetation  type  discrimination. 


BACKGROUND 

The  U.S.  Department  of  Agriculture,  Forest  Service,  has  developed  an 
increasing  interest  in  the  use  of  Advanced  Very  High  Resolution  Radiometer 
(AVHRR)  data  for  vegetation  assessments  of  broad  regions.  The  AVHRR  sensor 
is  mounted  on  National  Oceanic  and  Atmospheric  Administration  (NOAA)  polar- 
orbiting  weather  satellites. 

Typical  applications  of  AVHRR  data  for  forest  assessments  include  land  cover 
characterization  (Brown,  et  al.  1993;  Loveland  1991),  forest  density  analysis 
(Iverson,  et  al.  1989;  Ripple  1994;  Zhu  1994),  and  general  forest  distribution 

'  Approved  for  publication  as  Proceedings  Article  No.  FA-055-0296  of  the  Forest  and  Wildlife  Research  Center, 
Mississippi  State  University. 

2  Assistant  Professor  of  Forestry,  Mississippi  State  University,  Starkville,  MS. 

3  Mathematical  Statistician,  USDA  Forest  Service,  Rocky  Mountain  Forest  and  Range  Experiment  Station,  Fort  Collins, 
CO. 


419 


mapping  (Zhu  and  Evans  1994;  Evans,  et  al.  1992).  The  latter  of  these 
applications  utilized  both  the  original  spectral  data  and  Normalized  Difference 
Vegetation  Indices  (NDVTs)  in  various  interpretation  schemes  to  characterize 
earth  cover. 

The  accuracy  of  products  derived  from  AVHRR  data  is  difficult  to  assess  at 
continental  scales  due  to  the  comprehensive  coverage  of  the  data  and  the  low 
resolution  of  each  data  element  (1.1  km  at  nadir).  Therefore,  increased  attention 
has  been  devoted  to  use  of  data  analysis  techniques  that  optimize  chances  for  high 
classification  success.  One  possible  way  to  increase  classification  success  is  to 
identify  the  data  combinations  and  transformations  that  can  be  used  effectively  in 
automated  classification  procedures.  Exclusive  use  of  some  data  transformations 
could  deprive  classification  procedures  of  valuable  information  that  is  inherent  in 
the  original  data.  The  concern  identified  in  this  paper  is  that  NDVTs  (data 
transformations)  may  forfeit  useful  information  that  could  have  been  obtained  from 
the  original  AVHRR  spectral  channels. 

DERIVATION  AND  USES  OF  NDVTs 

The  NDVI  is  a  data  transformation  (Eq.l)  used  to  characterize  vegetation  vigor 
or  "greenness." 

NDVI  =  (NIR  -  VIS  /  NIR  +  VIS)  *  100  [1] 

In  Equation  [1],  NIR  is  the  reflectance  values  for  the  near-infrared  channel 
(channel  2  in  AVHRR)  and  VIS  is  the  reflectance  values  for  the  visible  channel 
(channel  1  in  AVHRR)  from  satellite  data. 

A  common  approach  used  to  interpret  AVHRR  data  is  to  acquire  imagery  over 
one  year's  time  and  calculate  NDVTs  to  document  vegetative  phenology.  This  is 
done  to  identify  different  or  distinctive  phenological  sequences  that  can  be 
attributed  to  vegetation  types,  crop  development,  and  regional  climate  conditions. 

NDVI  data  have  been  used  as  input  to  classification  algorithms  to  identify 
vegetative  cover  types  (Brown,  et  al.  1993;  Loveland,  et  al.  1991).  With  time  as  a 
critical  component  in  classification,  it  is  assuned  that  vegetative  cover  types  have 
different  phenological  time  frames.  Given  the  low  sensitivity  of  the  AVHRR 
instrument  to  spectral  variance  in  1.1  km  resolution  elements,  this  assumption  of 
separability  may  not  hold  in  the  classification  process. 

A  complicating  factor  in  vegetation  identification  is  that  the  same  vegetation  may 
exhibit  a  wide  range  of  latitudinal  variation  in  phenology.  There  are  numerous 
examples  of  wide-ranging  eastern  deciduous  trees  that  occur  from  the  deep  south 
to  northern  states.  Eastern  cottonwood  (Populus  deltoides,  Bartr.)  ranges  from 
southern  Louisiana  to  central  Minnesota  and  red  maple  {Acer  rubrum,  L.)  ranges 
from  central  Florida  to  southern  Canada.  Many  of  the  Oaks  (Quercus  spp.  )  and 
hickories  (Carya  spp.  )  also  exhibit  wide  geographic  ranges.     These  wide 


420 


geographic  ranges  can  easily  mean  that  leaf  fall  or  leaf  flush  could  vary  two  to 
three  months  across  the  range  for  any  given  species.  The  timing  of  phenologic 
events  for  any  given  cover  type  must  contribute  to  spectral  reflectance  pattern 
variation  that  can  ultimately  cause  confusion  in  image  classification. 

Confusion  between  different  cover  types  with  the  same  NDVTs  can  be  illustrated 
by  a  hypothetical  example.  Let  us  suppose  that  forest  land  in  a  region  has  spectral 
reflectance  values  of  20  and  40  for  the  visible  and  NIR  channels  in  AVHRR  data. 
Grassland  in  the  same  region  may  have  values  of  30  and  60  for  the  same  two 
channels.  Clearly  these  two  cover  types  are  distinguishable  in  the  separate  spectral 
channels  but  the  NDVTs  for  each  are  identical  (rounded  value  of  33).  In  this 
hypothetical  but  realistic  case,  the  use  of  NDVTs  inadvertently  filters  out  useful 
information  and  causes  confusion.  This  is  an  assumed  single-date  case  but  even  if 
there  were  multiple  dates  of  imagery,  it  is  possible  that  some  NDVTs  would  be  the 
same  for  otherwise  different  cover  classes  and  therefore,  would  degrade 
classification  performance. 

RESEARCH  OBJECTIVE 

The  basic  objective  of  use  of  AVHRR  data  by  the  USDA  Forest  Service  is  to 
distinguish  forest  cover  from  other  earth  cover  over  broad  continental-scale 
regions.  This  has  been  demonstrated  by  (Zhu  and  Evans  1994).  This  project 
utilized  NDVTs  and  the  original  spectral  data,  but  placed  significant  reliance  on 
ancillary  data  and  post-processing  techniques  to  resolve  conflicts  in  cover  class 
identification.  The  concern  for  future  uses  of  AVHRR  NDVTs  is  the  ability  to 
assess  earth  cover  in  regions  where  little  ancillary  information  is  available.  Lack  of 
ancillary  information  dictates  that  the  maximum  amount  of  information  has  to  be 
extracted  from  original  image  data.  This  project  was  initiated  to  evaluate 
information  content  in  satellite  data  to  better  define  appropriate  interpretation 
techniques  that  will  maximize  success  in  cover  type  identification. 

INITIAL  METHODOLOGY 

The  first  phase  of  this  project  is  to  compare  cover  class  separability  with  NDVTs 
versus  the  visible  and  NIR  AVHRR  channels  for  areas  of  known  ground  cover. 
Early  work  has  concentrated  on  use  of  data  from  an  on-going  project  to  map  the 
forest  cover  of  Mexico  and  Central  America  (Lannom  1995).  Landsat  TM  data  of 
Northeastern  Mexico  were  classified  as  agriculture  (generally  fallow  fields),  forest, 
and  scrub  land. 

AVHRR  data  used  in  this  initial  investigation  were  near-cloud-free  composites 
derived  for  North  America  by  the  maximum  NDVT  method  described  by 
Eidenshink  (1992).  These  data  sets  were  composed  of  information  compiled  over 
10-day  intervals.  Those  AVHRR  data  sets  that  did  not  exhibit  clouds  or 
substantial  compositing  boundaries  (due  to  abrupt  changes  in  reflectance  values 


421 


between  image  dates)  were  chosen  for  initial  study.  This  yielded  9  data  sets  from 
those  available  over  one  year  that  were  deemed  useful  for  the  study. 

The  TM  data  were  co-registered  to  the  AVHRR  data.  A  linked  display  system 
implemented  in  PCI's  Imageworks4  was  used  to  locate  3  areas  in  each  TM  cover 
class  that  covered  25  AVHRR  cells  (75  total  cells).  EASI/PACE  software  was 
then  used  to  export  the  AVHRR  visible,  NIR,  and  NDVI  data  values  to  a  text  file 
for  analysis  with  the  Splus  statistics  software  package.  The  discriminant  analysis 
function  in  Splus  was  used  to  generate  discriminant  correlations  (measure  of  the 
ability  to  identify  the  three  cover  classes)  for  the  visible,  NIR,  visible  and  NIR 
combined,  and  NDVI  channels  of  the  AVHRR  data  set. 

RESULTS  AND  DISCUSSION 

The  results  of  the  discriminant  analysis  work  are  presented  in  Table  1.  The 
statistical  significance  of  these  findings  has  not  been  determined  because  these  tests 
represent  a  limited  area  and  should  be  greatly  expanded  to  cover  a  wider  range  of 
conditions  that  are  assumed  to  occur  in  large  AVHRR  data  sets. 

Table  1. — Discriminant  correlations  and  averages  for  three  cover  classes  as  represented 
by  visible  NIR  and  NDVI  channels  from  selected  AVHRR  data  sets. 


AVHRR 

Channels 

Composite 

VIS 

NIR 

VIS  &  NIR 

NDVI 

Aprill 

0.9148 

0.5656 

0.9839 

0.9586 

April2 

0.6793 

0.4311 

0.9878 

0.8900 

April3 

0.9458 

0.6711 

0.9702 

0.9323 

June2 

0.9062 

0.2806 

0.9688 

0.9617 

July3 

0.9305 

0.2411 

0.9459 

0.8468 

Septl 

0.9549 

0.2290 

0.9734 

0.9445 

Sept2 

0.9297 

0.2883 

0.9773 

0.9187 

Oct2 

0.8337 

0.1872 

0.9470 

0.9175 

Novl 

0.8296 

0.1802 

0.9486 

0.8721 

Average 

0.8805 

0.3416 

0.9670 

0.9158 

It  was  interesting  to  note  that  on  average,  the  visible  channel  seemed  to  have 
better  discriminating  capabilities  for  the  three  subject  cover  classes  than  did  the 
NIR.  One  would  expect  that  the  NIR  channel,  which  is  sensitive  to  vegetation 
response,  would  provide  better  discriminating  capabilities.  This  area  of  Mexico 
(particularly  west  of  the  mountains)  is  dominated  by  fairly  arid  lands.  The 
agriculture  and  scrub  land  categories  undoubtedly  had  significant  amounts  of 
exposed  soil  that  influenced  the  visible  reflectance  channel  values.    The  visible 


Mention  of  company  or  product  names  is  for  information  only  and  does  not  constitute  official  endorsement  by  Mississippi 
State  University  or  the  USDA  Forest  Service. 


422 


channel  generally  exhibits  greater  contrast  and  variance  in  spectral  values  between 
vegetated  and  non-vegetated  cover.  Thus,  the  greater  contrast  in  visible  channel 
pixels  significantly  influenced  differentiation  of  forest  from  the  other  categories 
that  had  more  bare  ground  exposed  to  the  sensor. 

The  NDVI  channel  outperformed  the  visible  and  NCR  channels  when  used  singly. 
It  did  not  seem  to  discriminate  the  subject  classes  as  well  as  the  combined  visible 
and  NIR  channels.  This  finding  seems  to  support  the  original  premise  that  NDVI 
may  mask  useful  information  found  in  the  original  channels  that  could  be  used  to 
discriminate  different  vegetative  cover.  Additional  and  more  rigorous  testing  is 
recommended  to  substantiate  these  findings. 

FUTURE  WORK 

Additional  testing  using  these  or  modified  techniques  will  be  much  broader  based 
than  this  exploratory  study.  TM  and  AVHRR  data  sets  are  readily  available  for 
different  locations  in  Central  America,  Mexico,  and  the  U.S.  Testing  over  this 
broad  range  of  conditions  will  provide  a  better  determination  of  the  utility  of 
different  AVHHR  channels  in  vegetation  discrimination.  The  findings  from  this 
work  will  likely  have  broad  application  in  the  improvement  of  on-going  and  future 
studies  that  utilize  AVHRR  for  global  vegetation  distribution  assessments. 

LITERATURE  CITED 

Brown,  Jesslyn  F.,  Thomas  R.  Loveland,  James  W.  Merchant,  Bradley  C.  Reed, 
and  Donald  O.  Ohlen.  1993.  Using  multisource  data  in  global  land-cover 
characterization:  concepts,  requirements,  and  methods.  Photogramm.  Engr. 
and  Remote  Sensing.  59(6):977-987. 

Eidenshink,  Jeffery  C.  1992.  The  1990  conterminous  U.S.  AVHRR  data  set. 
Photogramm.  Engr.  and  Remote  Sensing.  58(6):809-813. 

Evans,  David  L.,  Zhiliang  Zhu,  Susan  Eggen-Mclntosh,  Pedro  Garcia  Mayoral, 
and  Jose  Luis  Ornelas  de  Anda.  1992.  Mapping  Mexico's  forest  lands  with 
advanced  very  high  resolution  radiometer.  USDA  Forest  Service,  Southern 
Forest  Experiment  Station,  Research  Note  SO-367,  4  p. 

Iverson,  L.R.,  E.A.  Cook,  and  R.  L.  Graham.  1989.  A  technique  for  extrapolating 
and  validating  forest  cover  across  large  regions:  calibrating  AVHRR  data  with 
TMdata.  International  J.  Remote  Sensing.  10(1 1):  1 805-1 8 12. 

Lannom,  Keith  B.  1995.  Mapping  forest  distributions  of  Mexico  and  Central 
America.  Geomatics  Info  Magazine.  9(5):33-37. 


423 


Loveland,  Thomas  R.,  James  W.  Merchant,  Donald  O.  Ohlen,  and  Jesslyn  F. 
Brown.  1991.  Development  of  a  land-cover  characteristics  database  for  the 
conterminous  U.S.  Photogramm.  Engr.  and  Remote  Sensing.  57(11):  1453- 
1463. 

Ripple,  William  J.  1994.  Determining  coniferous  forest  cover  and  forest 
fragmentation  with  NOAA-9  very  high  resolution  radiometer  data. 
Photogramm.  Engr.  and  Remote  Sensing.  60(5):533-540. 

Zhu,  Zhiliang.  1994.  Forest  density  mapping  in  the  lower  48  States,  a  regression 
procedure.  USDA  Forest  Service,  Southern  Forest  Experiment  Station, 
Research  Paper  SO-280.  1 1  p. 

Zhu,  Zhiliang,  and  David  L.  Evans.  1994.  U.S.  forest  types  and  predicted  percent 
forest  cover  from  AVHRR  data.  Photogramm.  Engr.  and  Remote  Sensing. 
60(5):525-531. 

BIOGRAPHICAL  SKETCHES 

David  L.  Evans  is  an  Assistant  Professor  of  Forestry  at  Mississippi  State 
University,  Starlcville,  MS.  He  completed  the  Ph.D.  in  Forest  Management  at 
Louisiana  State  University  in  1986.  His  teaching  and  research  interests  are  in  the 
use  of  remote  sensing  and  geographic  information  systems  technologies  for  natural 
resource  assessments.  Dr.  Evans  was  formerly  with  the  USDA  Forest  Service  as  a 
Supervisory  Research  Forester  and  Team  Leader  for  remote  sensing  research  in 
the  Southern  Research  Station. 

Ray  Czaplewski  has  spent  the  past  13  years  working  on  statistical  design  and 
analysis  of  regional  monitoring  systems  for  forests  and  other  environmental 
features,  primarily  the  Forest  Service's  nationwide  Forest  Inventory  and  Analysis 
Program.  Other  accomplishments  include  statistical  design  of  the  1990  and  2000 
global  surveys  of  forests  by  the  Food  and  Agricultural  Organization  of  the  United 
Nations,  and  design  to  monitor  the  effectiveness  of  the  President's  1994  Forest 
Plan  for  the  States  of  Oregon,  Washington,  and  California.  He  has  specialized  in 
linking  sample  surveys  to  remotely  sensed  data  and  geographic  information 
systems. 


424 


Measures  of  Association  and 
Agreement  for  Describing  Land  Cover 
Characterization  Classes 

Eugene  A.  Fosnight1  and  Gary  W.  Fowler2 


Abstract.-  Several  statistical  measures  of  association  or  agreement  are 
available  for  comparing  land  cover  characterizations  to  reference  data. 
This  study  compares  five  statistical  measures  under  three  sampling 
schemes  for  five  different  class  aggregations.  Results  show  that  each 
statistical  measure  contributes  to  the  overall  view  of  the  information. 
Aggregated  land  cover  characterization  classes  are  less  successful  in 
predicting  the  reference  classes.  Sample  design  and  size  can  cause  biased 
estimators. 


INTRODUCTION 

Confidence  in  land  cover  characterizations  derived  from  remotely  sensed 
images  is  acquired  through  a  thorough  description  of  the  characterizations.  This 
paper  seeks  to  describe  measures  of  agreement  and  association  suitable  for 
quantifying  the  relationship  between  ground- sampled  or  photointerpreted  data  and 
characterizations  of  remotely  sensed  images.  This  objective  consists  of  three  tasks: 
(1)  describe  a  subset  of  measures,  (2)  analyze  the  distribution  of  each  measure 
under  one  simple  and  two  stratified  random  sampling  schemes,  and  (3)  discuss 
issues  of  interpretability  associated  with  each  statistic. 

For  this  study  the  U.S.  Land  Cover  Characterization  database  was  compared  to 
a  photointerpreted  land  use  and  land  cover  data  set.  The  measures  of  association 
and  agreement  calculated  were  Cohen's  Kappa,  Cramer's  V,  Guttman's 
Proportional  Reduction  in  Error  (PRE),  Goodman  and  Kruskal's  Proportion  of 
Explained  Variance  (PEV),  and  Percent  Correctly  Classified  (PCC). 

MEASURES  OF  ASSOCIATION  AND  AGREEMENT 

Many  measures  of  association  and  agreement  are  available  for  quantifying  the 
cross-classified  tables  that  are  created  by  cross-tabulating  classified  images  and 
reference  data.  Within  the  remote  sensing  literature,  these  tables  are  often  referred  to 
as  "confusion  tables"  or  "error  matrices"  (Czaplewski  and  Catts  1992,  Story  and 
Congalton  1986). 

A  few  general  concepts  should  be  defined.  The  probability  for  row  i  and 
column  j  is  pt].  The  estimated  value  of  ptj  is  ptj  =xij/N  where  xtj  is  the  cell 


'Cartographer,  Hughes  STX,  UNEP/GRID.  EROS  Data  Center,  Sioux  Falls,  SD.  Work  performed  under  U.S. 
Geological  Survey  contract  1434-92-C-40004. 

Biometrician,  School  of  Natural  Resources  and  Environment,  University  of  Michigan,  Ann  Arbor,  MI. 


425 


count.  The  expected  value,  m- ,  of  xi}  given  no  information  other  than  the  sample 
marginal  frequencies,  xi+  and  x+j  is  mV]  =  pi+p+j ,  where  pi+  are  the  row  marginal 
probabihties  and  p+j  are  the  column  marginal  probabilities. 

Measures  based  on  Chi-Square 

Several  measures  of  association  are  derivatives  of  Pearson's  Coefficient  <I>2  of 
Mean-Square  Contingency  (Eq  1).  Cramer  suggested  modifications  to  restrict  the 
range  to  between  0  and  1  (Bishop  et  al.  1975,  Conover  1980).  The  statistical 
measures  are 

(£Li^and  Cramer's  V  =  (  £  f,  [1] 

p^p^  [min[(/-l),(/-l)]J 

where  the  maximum  achievable  value  of  O  is  min[(7  - 1),  (/  - 1)] . 

As  V  depart  from  zero,  the  measures  are  interpreted  as  becoming  less 
independent. 

Proportional  Reduction  in  Error 

Goodman  and  Kruskal  (1979)  discussed  the  PRE  (A)  measure  based  on  a 
probabilistic  model  (Eq.  2).  Two  probabihties  are  compared  in  calculating  PRE. 
First,  the  probability  of  an  error  in  correctly  guessing  the  column  by  chance  is 
determined  by  calculating  the  most  frequently  occurring  column.  This  probability  is 

1 "  P+m .  where  p+m  =  max(/?+1 , p+2 ,. . . , p+J ) . 

Second,  the  probability  of  correctly  guessing  the  column  given  rows  is 
determined  by  calculating  which  rows  occur  most  frequently  with  each  column. 
This  probability  is 

1  -  X  Pim  >  where  Pim  =  max(p/1 ,  pi2 , . . . ,  pa  ) . 
1 

/  / 

(i-P+J-a-IX,)  ^  ^xim-x+m 

XaR  =  !   and  XCiR  =  [2] 

N-x+m 

If  the  variables  are  statistically  independent,  PRE  will  be  zero;  however,  if  PRE 
is  zero,  the  variables  are  not  necessarily  statistically  independent.  The  predictive 
interpretation  of  PRE  distinguishes  it  from  Chi-square-based  measures. 

Proportion  of  Explained  Variance 

After  Bishop  et  al.  (1975),  Guttman's  PEV  (  r)  can  be  explained  as  the 
difference  between  predicting  the  row  given  only  the  marginal  probabilities  of  the 
columns  and  predicting  the  column  given  the  conditional  proportions  of  the  rows 
(Eq.  3). 

PEV  can  be  cast  into  analysis  of  variance  terminology,  where  the  total  variation 
is  defined  as: 

2     2Nft  +J 


426 


This  definition  of  total  variation  meets  the  criterion  of  equaling  zero  if  and 
only  if  all  counts  are  in  the  same  column,  and  it  is  maximized  when  the  marginal 
frequencies  of  the  columns  are  uniformly  distributed. 

The  total  variation  is  split  into  between  (BSS)  and  within  (WSS)  class 
components 

WSS  =  Y 4£t-2W  -and  BSS  =  TSS-WSS  =  ii±i4-±-iXli. 

PEV  is  then  set  equal  to  the  ratio  of  the  between  class  component  and  the  total 
variation. 

*c\r  TSS- 

The  measure  PEV  for  the  column  given  the  rows  can  be  described  as  the 
percentage  of  the  variation  in  the  columns  that  can  be  explained  knowing  the  rows. 

Measures  of  Agreement 

Measures  of  agreement  have  more  restrictive  class  specifications  than  do 
measures  of  association,  but  they  also  provide  more  powerful  tests  if  the  class 
specifications  can  be  met.  The  cross-classified  table  must  have  identical  classes  in 
the  same  order  for  both  rows  and  columns.  This  restriction  gives  special  meaning 
to  the  diagonal  of  the  table. 

PCC,  the  simplest  measure  of  agreement,  is  calculated  by  dividing  the  sum  of 
the  diagonal  cells  (all  samples  correctly  classified)  by  the  total  number  of  samples 
(Eq.  4). 

PCC  =  £ Pii  and  PCC  =  £ xu /  /V .  [4] 

;=i  1=1 

This  is  the  most  common  statistic  for  assessing  accuracy  in  the  remote  sensing 

literature.  PCC  has  a  highly  intuitive  interpretation. 

Cohen  (1960)  proposed  a  measure  of  agreement,  Kappa,  (Eq.  5)  that  adjusts 

PCC  for  the  probability  of  chance  agreement.  Chance  agreement  is  the  probability 

of  agreement  given  only  the  sample  marginal  proportions. 

i  i  i  i  ii 

X ft "  X'7z«  Eft  N%xu  -  X*,+*+i 

K  =  =  i=l  pi^  and  K  =  — ^  ^   [5] 

i  -  X  m«      1 "  X  Pi+p+i  Nl  -  X  xi+x+i 

1=1  1=1  1=1 

Kappa  equals  1  if  there  are  no  off-diagonal  counts.  With  the  exception  of 
perfect  agreement  Kappa  will  always  be  lower  than  PCC.  Unfortunately,  as  noted 
by  Card  (1982)  and  Hay  (1979),  PCC  can  also  be  a  highly  biased  estimator. 

Calibration 

The  estimated  probability  of  each  cell  in  the  cross-classified  table  (Eq.  6)  can  be 

calibrated  by  incorporating  the  known  population  marginal  frequencies: 

*     X  p  J 
Pii  ~    1  U »  wnere  X  ■  =  Y,^  are  tne  population  marginal  frequencies.  [6] 
J     Np  ;  % 


427 


Calibrated  estimates  of  the  cell  counts  (Eq.  7)  can  then  be  calculated  from  the 
cell  probabilities  and  the  known  finite  population  is 

x*j  =  P*jN.  [7] 

Card  (1982)  and  Czaplewski  and  Catts  (1992)  described  in  detail  calibration 
techniques  for  deriving  calibrated  estimates  for  the  areas  of  reference  classes  given 
known  population  frequencies  of  the  modeled  classes. 

The  measures  of  association  and  agreement,  Cramer's  V,  Guttman's  PRE, 
Goodman  and  Kruskal's  PEV,  PCC,  and  Cohen's  Kappa  form  a  representative 
sample  of  the  measures  that  could  contribute  toward  assessing  the  accuracy  of 
remotely  sensed  classifications. 

EXPERIMENTAL  DATA 

The  study  area  is  the  conterminous  United  States.  The  U.S.  Land  Cover 
Characterization  database  is  compared  to  the  reference  U.S.  Geological  Survey 
(USGS)  Land  Cover  and  Land  Use  digital  data.  The  U.S.  Land  Cover 
Characterization  database  is  derived  from  Advanced  Very  High  Resolution 
Radiometer  (AVHRR)  images  (Loveland  et  al.  1991,  Brown  et  al.  1993).  The 
USGS  Land  Cover  and  Land  Use  digital  data  are  derived  from  photographs  at 
scales  smaller  than  1:60,000  (Anderson  1976,  USGS  1990). 

The  two  classifications  were  grouped  into  several  alternative  aggregations. 
Three  levels  of  the  Anderson  classification  scheme  were  created  for  the  reference 
data  set:  (1)  the  37-class  Anderson  level  II  (lulc),  (2)  a  14-class  modified  Anderson 
that  retains  the  classes  of  natural  vegetation,  range  (mixed,  shrub  and  brush,  and 
herbaceous),  forest  (deciduous,  evergreen,  and  mixed),  and  wetlands  (forested  and 
nonforested)  at  Anderson  Level  II,  while  collapsing  the  remaining  classes  to 
Anderson  Level  I  (lulcmA),  and  (3)  an  8-class  Anderson  Level  I  (lulcl). 

The  level  I  classification  was  determined  by  Anderson  to  be  the  class  resolution 
attainable  by  Landsat  Multispectral  Scanner  scale  classifications.  The  Level  II 
classification  was  considered  attainable  using  small-scale  (less  than  1:60,000)  aerial 
photographs.  The  detailed  temporal  information  available  with  AVHRR  images 
(daily  coverage)  and  the  ancillary  information  provide  the  possibility  of  attaining  at 
least  level  I  class  resolutions. 

Three  classifications  were  created  from  the  land  cover  characterization:  (1)  the 
ungrouped  classes  (lcc),  (2)  an  aggregation  corresponding  to  the  lulcl  (led)  classes, 
and  (3)  an  aggregation  corresponding  to  the  lulcmA  (lccmA)  classes. 

The  classifications  of  the  land  cover  characterization  classes  and  reference  land 
use  land  cover  classes  were  cross-classified  into  five  tables;  lcclulc,  lcclulcmA, 
lcclulcl,  lccmAlulcmA,  lccllulcl.  The  first  three  cross-classified  tables  (lcclulc, 
lcclulcmA,  and  lcclulcl)  are  designed  to  show  how  well  the  full  set  of  land  cover 
characterization  classes  can  predict  decreasing  levels  of  class  resolution  within  the 
Anderson  classification.  The  square  cross-classified  tables  use  Cohen's  Kappa  and 
PCC  statistics  for  measuring  agreement  between  the  variables. 


428 


RESULTS 


The  parameters  provide  a  baseline  for  comparison  with  simulation  results. 
When  lcc  is  compared  to  the  three  lulc  aggregations,  the  ability  of  lcc  to  predict  lulc, 
as  measured  by  PREiuicucc,  increases  as  the  number  of  classes  in  lulc  decreases, 

likewise,  the  proportion  of  explained  variance,  as  measured  by  PEViu]cucc,  and  the 

lack  of  independence,  as  measured  by  Cramer's  V,  also  increase.  The  least  amount 
of  predictive  or  explanatory  power  exists  when  lcc  is  aggregated  to  the  modified 
Anderson  classification,  where  the  class  resolution  of  the  land  cover 
characterization  classes  are  stretched  to  the  limit.  When  the  data  sets  are  aggregated 
to  class  resolution  more  appropriate  to  the  AVHRR  satellite,  as  in  lccilulcl,  the 
measures  approach  their  maximum. 

The  results  of  the  Monte  Carlo  simulations  for  simple  random  sampling, 
stratified  random  sampling  with  probability  proportional  to  size,  and  stratified 
random  sampling  with  equal  probabilities  are  shown  in  Figures  1,  2,  and  3, 
respectively.  The  Monte  Carlo  simulations  consisted  of  a  thousand  repetitions 
calculated  for  each  of  the  measures  for  random  samples  selected  without 
replacement  from  the  full  data  set.  The  population  parameters  are  shown  on  each  of 
the  figures. 

All  of  the  measures  have  inflated  values  for  small  sample  sizes  extracted  from 
the  large  rectangular  tables.  Whether  this  is  caused  by  the  large  number  of  cells 
(and  large  number  of  empty  cells)  or  the  rectangular  nature  of  the  tables  is 
unknown;  however,  the  effect  is  least  pronounced  in  table  lcclulcl,  which  is  the 
most  rectangular  and  has  the  fewest  empty  cells.  Cramer's  V  shows  some  bias  for 
all  tables  at  small  sample  sizes. 

For  simple  random  sampling  and  stratified  random  sampling  with  probability 
proportional  to  size,  the  distributions  are  very  similar.  Stratified  random  sampling 
with  probability  proportional  to  size  shows  slightly  larger  biases  for  small  sample 
sizes.  The  stratified  sampling  by  lcc  class  enforced  a  more  even  distribution  of 
samples  across  lulc  classes,  causing  fewer  classes  with  a  marginal  frequency  of 
zero. 

Under  stratified  random  sampling  with  equal  probabilities,  PRE,  PEV,  and 
Cramer's  V  are  consistently  overestimated,  while  for  Kappa  and  PCC  the  measures 
are  consistently  underestimated  (Figure  3). 

Equal  probability  sampling  results  in  small  classes  being  over-sampled  and 
large  classes  being  under-sampled.  For  large  percent  samples,  the  small  classes  are 
exhaustively  sampled,  and  hence  cannot  achieve  the  target  percentage.  Certainly  in 
the  case  of  a  10  percent  sample,  all  159  lcc  classes  cannot  be  sampled  at  10  percent 
of  the  total.  The  very  smallest  classes  are  nearly  exhaustively  sampled  with  the  .0 1 
percent  sample. 

Because  equal  probability  sampling  is  not  necessarily  representative  of  the 
population,  it  becomes  susceptible  to  differential  accuracy  among  the  classes.  The 
net  result  is  that  the  statistical  measures  are  highly  biased  for  even  large  sample 
sizes,  and  the  direction  of  the  bias  depends  on  the  statistic  and  per  class  accuracies. 


429 


PRE  clr 


PEV  clr 


V 


Icclulc 


IcclulcmA  § 


Icclulcl 


IccmAlulcmA 


Iccllulcl 


b  a 


ft  B 


o 
o 

 — — 9  a 

o 

B — ill 

M 

.01%.1%  1%  10% 


6  a 


Kappa 


-B — B- 


.0I%.1%  1%  10% 


.01%.1%  1%  10% 


.01%.1%  1%  10% 


PCC 


$  i  a 


.0I%.1%  1%  10% 


Figure  l.-Results  from  Monte  Carlo  simulation  of  simple  random  sample  for  measures 
of  association  and  agreement.  Missing  boxplots  result  either  from  all  zero  cells  for  an 
lcc  or  an  lulc  class  or  from  computation  time  too  long  to  complete.  The  horizontal  lines 
through  the  boxes  are  the  population  parameters. 


Icclulc 


IcclulcmA 


Icclulcl 


IccmAlulcmA 


Iccllulcl 


PRE  clr 


?  SB 


-5-s- 


.01%.1%  1%  10% 


PEV  clr 


T 

s — g- 


.01%.1%  1%  10% 


V 


T 

-n — 5- 


Kappa 


PCC 


-b — a- 


T 

-B — S" 


.01%.1%  1%  10% 


01%.1%  1%  10% 


.01%.1%  1%  10% 


Figure  2.-Results  from  Monte  Carlo  simulation  of  stratified  random  sampling  with 
probability  proportional  to  size  for  measures  of  association  and  agreement.  Missing 
boxplots  result  from  all  zero  cells  for  either  an  lcc  or  an  lulc  class.  The  horizontal  lines 
through  the  boxes  are  the  population  parameters. 


430 


PRE  clr 


PEVclr 


V 


Icclulc 


IcclulcmA 


Icclulcl 


IccmAlulcmA 


2!  6 


Kappa 


PCC 


Iccllulcl 


a  ffi 


□  s 


.01%. 1%  1%  10% 


.01%. 1%  1%  10% 


.01%. 1%  1%  10% 


.01%. 1%  1%  10% 


01%    1%    1%  10% 


Figure  3.-Boxplots  of  results  from  Monte  Carlo  simulation  of  stratified  random 
sampling  with  equal  probability  allocation  for  measures  of  association  and  agreement. 
Missing  boxplots  result  from  all  zero  cells  for  either  an  Ice  or  an  lulc  class.  The 
horizontal  lines  through  the  boxes  are  the  population  parameters. 

Equal  probability  sampling  ensures  a  sufficient  sample  size  of  each  stratified 
class.  This  allows  a  sufficient  description  of  each  of  the  stratified,  in  this  case  lcc, 
classes.  Under  probability  proportional  to  size,  sampling  some  classes  may  be 
under-represented.  For  example,  rare  classes  may  be  sampled  only  once,  if  at  all, 
to  meet  strict  probability  proportional  to  size  criteria. 

The  cell  counts  resulting  from  stratified  random  sampling  with  equal 
probabilities  can  be  adjusted  to  better  conform  to  cell  probabilities  yielded  through 
simple  random  sampling.  In  this  Monte  Carlo  simulation,  each  sample  from  the 
cross-classified  tables  Icclulcl  and  Iccllulcl  was  selected  with  equal  probabilities  and 
was  calibrated  using  the  population  marginal  sums  for  the  lcc  classes.  The  results 
from  the  calibration  are  shown  in  Figure  4. 


PRE  clr 


PEVclr 


V 


Icclulcl 


Iccllulcl 


m  3 


Kappa 


PCC 


-B  ffi- 


-s — »- 


01%.l%  1%  10% 


.01%. 1%  1%  10% 


.01%. 1%  1%  10% 


.01%. 1%  1%  10% 


.01%. 1%  1%  10% 


Figure  4.-Results  from  Monte  Carlo  simulation  of  stratified  random  sampling  with 
equal  probability  allocation  for  measures  of  association  and  agreement  after  calibration 
of  cell  counts.  The  horizontal  lines  through  the  boxes  are  the  population  parameters. 

The  distributions  after  calibration  for  the  cross-classified  table  Iccllulcl  are 
very  similar  to  those  from  the  simple  random  sample.  For  the  larger  rectangular 
table  Icclulcl,  even  though  the  biases  are  removed  for  large  sample  sizes,  the  biases 


431 


for  small  samples  grew  to  be  even  larger.  For  this  type  of  table,  the  small  sample 
sizes  and  large  number  of  empty  cells  prevent  the  calibration  from  being  well 
controlled.  Unfortunately,  these  are  the  practical  sample  sizes  for  actual 
applications. 

CONCLUSION 

Only  three  of  the  measures,  PRE,  PEV,  and  Cramer's  Y,  are  suitable  for 
rectangular  tables.  PRE  and  PEV  are  asymmetrical  predictive  measures,  and 
Cramer's  V  is  a  measure  of  independence.  Cramer's  V  shows  some  tendency  to  be 
biased  for  small  sample  sizes,  even  for  a  small  number  of  classes. 

Both  PEV  and  PRE  were  designed  for  ease  of  interpretation.  Cramer's  V  can 
perhaps  be  used  best  in  tandem  with  other  measures.  For  example,  a  table  can  be 
tested  for  independence  with  Cramer's  V,  and  then  the  strength  of  the  association 
can  be  measured  with  PRE  and  PEV.  For  those  cross-classified  tables  where 
Kappa  is  applicable,  it  provides  the  most  robust  statistic;  however,  PCC  is  the 
simplest  to  understand. 

REFERENCES 

Anderson,  J.R.,  Hardy,  E.E.,  Roach,  J.T.,  and  Witmer,  R.E.,  1976,  A  Land  Use 

and  Land  Cover  Classification  System  for  Use  with  Remote  Sensor  Data: 

Washington,  DC,  U.S.  Government  Printing  Office. 
Bishop,  Y.M.M.,  Fienberg,  S.E.,  and  Holland,  P.W.,  1975,  Discrete  Multivariate 

Analysis:  Cambridge,  Massachusetts,  The  MIT  Press. 
Brown,  J.F.,  Loveland,  T.R.,  Merchant,  J.W.,  Reed,  B.C.,  and  Ohlen,  D.O.,  1993, 

Using  Multisource  Data  in  Global  Land-Cover  Characterization:  Concepts, 

Requirements,  and  Methods:  Photogrammetric  Engineering  and  Remote 

Sensing,  v.  59,  no.  6,  p.  977-987. 
Card,  D.H.,  1982,  Using  Known  Map  Category  Marginal  Frequencies  to  Improve 

Estimates  of  Thematic  Map  Accuracy:  Photogrammetric  Engineering  and 

Remote  Sensing,  v.  48,  no.  3,  p.  431-439. 
Cohen,  J.,  1960,  A  Coefficient  of  Agreement  for  Nominal  Scales:  Educational  and 

Psychological  Measurement,  v.  20,  no.  1,  p.  37-46. 
Conover,  W.J.,  1980,  Practical  Nonparametric  Statistics:  New  York,  NY:  John 

Wiley  &  Sons. 

Czaplewski,  R.L.,  and  Catts,  G.P.,  1992.  Calibration  of  Remotely  Sensed 

Proportion  or  Area  Estimates  for  Misclassification  Error:  Remote  Sensing  of 

Environment,  v.  39,  pp.  29-43. 
Goodman,  L.A.,  and  Kruskal,  W.H.,  1979,  Measures  of  Association  for  Cross 

Classifications,  Springer  Series  in  Statistics:  New  York,  NY:  Springer- 

Verlag. 

Hay,  A.M.,  1979,  Sampling  Designs  to  Test  Land-Use  Map  Accuracy: 
Photogrammetric  Engineering  and  Remote  Sensing,  v.  45,  no.  4,  p.  529-533. 

Loveland,  T.R.,  Merchant,  J.W.,  Ohlen,  D.O.,  and  Brown,  J.F.,  1991, 
Development  of  a  Land-Cover  Characteristics  Database  for  the  Conterminous 


432 


U.S.:  Photogrammetric  Engineering  and  Remote  Sensing,  v.  57,  no.  11,  p. 
1,453-1,463. 

Story,  M.,  and  Congalton,  R.G.,  1986,  Accuracy  Assessment:  A  User's 
Perspective:  Photogrammetric  Engineering  and  Remote  Sensing,  v.  52,  no.  3, 
p.  397-399. 

U.S.  Geological  Survey,  1990,  Land  Use  and  Land  Cover  Digital  Data  from 
1:250,000-  and  l:100,000-Scale  Maps:  Reston,  Virginia,  U.S.  Geological 
Survey. 

ACKNOWLEDGMENTS 

The  research  described  in  this  article  has  been  supported  by  the  U.S. 
Environmental  Protection  Agency  (EPA)  through  Interagency  Agreement  IAG 
DW14936073  to  the  U.S.  Geological  Survey  and  through  the  National  Aeronautics 
and  Space  Administration  in  support  of  the  United  Nations  Environment 
Programme.  However,  it  has  not  been  subjected  to  EPA  review  and  therefore  does 
not  necessarily  reflect  the  views  of  the  agency.  No  official  endorsement  should  be 
inferred  by  either  the  U.S.  Government  or  the  United  Nations  Environment 
Programme. 

BIOGRAPHICAL  SKETCH 

Eugene  A.  Fosnight  is  a  cartographer  with  the  United  Nations  Environmental 
Programme's  Global  Resource  Information  Database  (GRID),  Sioux  Falls,  S. 
Dak.  He  graduated  from  Purdue  University  with  a  B.S.  in  1972,  from  University 
College  Swansea  with  a  diploma  in  cartography  in  1980  and  from  University  of 
Michigan  with  a  M.S.  in  remote  sensing  in  1992.  Gene  provides  cartographic, 
remote-sensing  and  statistical  support  for  GRID-Sioux  Falls. 

Gary  W.  Fowler  is  a  Professor  of  biometrics  with  the  School  of  Natural 
Resources  and  Environment  at  the  University  of  Michigan,  Ann  Arbor,  Mich.  He 
graduated  from  the  University  of  California  with  a  Ph.D.  in  1969.  Gary's  research 
concentrates  on  sequential  sampling,  estimation  of  forest  stem  parameters,  volume- 
basal  area  ratios  for  use  in  horizontal  point  sampling,  efficient  sampling  of  endemic 
forest  insect  populations,  development  of  new  volume  equations  for  commercial 
coniferous  species,  and  statistical  properties  of  species  diversities. 


433 


Forest  Spatial  Surveys  Using  the  Rao- 
Hartley-Cochran  Sampling  Design 

Jeffrey  S.  Pontius1 


Abstract. -An  illustration  of  the  Rao-Hartley-Cochran  (RHC)  sampling 
strategy  for  a  simulated  forest  inventory  is  presented.  The  objective  of 
this  illustration  is  to  promote  the  consideration  of  the  RHC  strategy  when 
a  probability  proportional  to  size  with  replacement  strategy  is  being 
considered  in  a  forest  inventory.  The  illustration  is  based  on  a  sampling 
frame  of  partitioned  selection  areas  of  the  forest  floor,  and  comparing  the 
RHC  strategy  to  a  Hansen-Hurwitz  probability  proportional  to  size 
strategy. 


INTRODUCTION 

This  note  presents  an  illustration  of  the  Rao-Hartley-Cochran  sampling  strategy 
in  the  context  of  forest  sampling  when  the  selection  units  are  partitioned  areas  of  a 
forest  floor.  In  forest  surveys,  the  goal  is  to  obtain  reliable  estimates  of 
characteristics  of  a  forest  based  on  a  sampling  strategy.  Naturally,  one  would  like 
to  use  a  sampling  strategy  that  results  in  unbiased  estimates  and  'small'  standard 
errors  of  those  estimates.  The  Rao-Hartley-Cochran  strategy  can  provide  smaller 
standard  error  estimates  as  compared  to  a  with  replacement,  probability 
proportional  to  size  strategy. 

Sampling  Frame 

There  are  various  approaches  to  sampling  a  forest  based  on  particular  sampling 
frames  (de  Vries  1989),  which  depend  on  the  objectives  of  the  survey.  The 
sampling  frame  used  in  this  illustration  is  based  on  Roesch,  Green  and  Scott 
(1993).  This  sampling  frame  is  constructed  by  partitioning  the  forest  floor  into 
mutually  exclusive  segments  based  on  the  (overlapping)  trees'  K-circles  (a  circle 
centered  at  the  tree  with  a  fixed  radius).  The  resulting  segments  are  then  selected, 
using  some  randomization  mechanism,  with  probabilities  proportional  to  the  size  of 
the  segments.  The  trees  whose  selection  areas  contain  the  randomly  selected 
segment  are  included  in  the  sample. 

Following  the  notation  in  Roesch,  Green  and  Scott  (1993),  let  {(1,  y{),  (2,  y2X 
(N,  yN)}  denote  the  finite  population  of  trees  with  their  unit  labels  and 


Assistant  Professor,  Department  of  Statistics,  Kansas  State  University,  Manhattan,  KS. 


434 


associated  values  of  a  characteristic  of  interest.  Let  \A\,  Ai,  As  }  be  the  set  of 
selection  areas  of  the  trees  formed  by  their  K-circles.   The  set  of  segment  areas 

{Aj,  j  =  1,  2,  M},  determined  by  the  overlapping  areas,  A,,  or  Aj  =  A,  for 
some  nonoverlapping  selection  areas,  forms  the  sampling  frame. 

A  REV  IEW  OF  THE  SAMPLING  STRATEGIES 


A  brief  review  of  the  Hansen-Hurwitz  and  the  Rao-Hartley-Cochran  strategies 
using  the  above  sampling  frame  is  presented. 

Hansen-Hurwitz  Estimation 

Roesch,  Green  and  Scott  (1993)  constructed  a  sampling  strategy  that 
incorporated  Hansen-Hurwitz  (1943)  {HHj  estimation.     The  probability  that 

selected  segment  j  adds  tree  i  to  the  sample  is  p;j  =  (Aj  I  A;)Zij, 

where  Z  =  s    '  seSment  3  ls  Part  °f *ne  ?th  tree  s  K-circle 
y     \  0,  otherwise 

N 

and  the  value  associated  with  segment  j  is  y}  =  52  Vij  Hi-  An  estimator  of  the 

M  N  _ 

parametric  total  Y  =  52y, ■=  52y, is 


1     m  77 


where  p-  =  AjIAj  and  Aj  is  the  area  of  the  ground  in  the  forest  to  be  surveyed. 
The  variance  of  the  estimator  in  Eq.  1  is 

with  variance  estimator 


1       m  (     v  \ 

=  —  ~T  AT-j-  -  Ymi  J 

m(m-\)pf\     Aj  / 


435 


Rao-Hartley-Cochran  Estimator 


Estimators  based  on  without  replacement  sampling  are  generally  more  efficient 
(i.e.,  the  estimators  have  smaller  variance)  than  estimators  based  on  with 
replacement  sampling.  The  well-known  Rao-Hartley-Cochran  sampling  strategy 
(Rao,  Hartley  and  Cochran  1962)  {RHC}  applied  to  the  segmentation  sampling 
frame  is  as  follows.  Partition  the  universe  U  of  M  segments  into  a  random 
grouping  {Qi,  Q2,      Qg,      Gm},  where  1  <  Mg,  the  groups  are  selected  with 

m 

equal  probabilities,  and  M  =  52  Mg.  Simple  random  sampling  without  replacement 

<ri 

(SRSWOR)  can  be  used  to  assign  segments  to  groups  by  selecting  a  SRSWOR  of 
Mi  segments  from  M  (for  group  Q\),  then  selecting  a  SRSWOR  of  M2  segments 
from  U—Gu  then  selecting  a  SRSWOR  of  M3  segments  from  U—  (ftUfc), 
and,  finally,  assigning  the  remaining  segments  to  Qm.  Then,  independently  from 
each  group,  select  one  segment  with  probability  proportional  to  the  normed  size 
measures  Aj  within  each  group.  Specifically,  those  probabilities  are  pg  =  Pj/Pg,  g 
=  1,2,...,  Mg,  for  segment  j  in  group  Qg,  where  pj  =  Aj/At  and  Pg  =  52  Pj- 

Let  yg  and  Ag  be  those  m  values  of  y3  and  Aj,  respectively,  selected  from  group 
g.  The  unbiased  RHC  estimator  of  the  parametric  total  Y  is 

Fkhc  =  At^P^-  (3) 

Chauduri  and  Mitra  (1992)  indicated  that  the  variance  derived  by  Rao,  Hartley  and 
Cochran  (1962)  usually  performs  as  well  as  an  alternative  variance  estimator 
proposed  by  Ohlsson  (1989),  so  the  RHC  variance  estimator  was  used  in  the 
simulation  illustration.  Rao,  Hartley  and  Cochran  (1962)  gave  several  group  size 

A  A 

conditions  for  comparing  the  variance  of  Frhc  to  the  variance  of  I^hh  In  the 
illustration  (see  next  section),  the  variance  minimizing  condition  M\  =  M2  =  ...  = 
Mm  =  M/m  applies,  so  the  form 

(  m-\\ 

V(Y^c)=\Kl-j^JV(Ymi), 

with  variance  estimator 


436 


1      (M-m\  "  /      y,      *  \2 

^-^{^)p/*t-r-'J  (4) 

was  used.  Note  that  there  has  been  some  discussion  about  how  well  the  efficiency 
of  the  RHC  strategy  holds  as  compared  to  some  with  replacement  sampling 
strategies  (Deshpande  1984,  Singh  and  Kishmore  1975).  The  simulation 
illustration  provides  an  additional  view  on  this  issue. 


AN  ILLUSTRATIVE  COMPARISON  OF  THE  STRATEGIES 


A  small  simulation  study  was  designed  to  investigate  some  of  the  empirical 
properties  of  the  RHC  and  the  HH  strategies  for  the  above  segmentation  sampling 
frame.  A  dispersion  pattern  of  N  =  100  "trees"  was  obtained  by  generating  the 
coordinates  (x,  y)  using  independent  random  values  from  Poisson(60)  + 
Uniform(0,  1)  (figure  1).  The  volume  (m3)  of  each  tree  was  generated  using 
independent  normal(8,  2)  random  values  based  on  the  timber  volumes  in  an 
example  of  de  Vries  (1989,  p.  77).  Note  that  the  use  of  the  Poisson  and  normal 
distributions  give  'well-behaved'  characteristics  to  the  data.  The  segmentation 
sampling  frame  was  laid  out  as  described  above,  the  M  =  156  areas  Aj  computed, 
and  the  pairs  (Aj ,  y  -),  entered  into  a  data  list.  Here,  At  is  the  sum  of  the  Aj . 
Using  the  Mlm  criterion  with  M  =  156,  the  following  set  of  nontrivial  pairs) (m, 
Mg)}  was  available  for  RHC  sampling  {(2,  78),  (4,  39),  (6,  26),  (12,  13),  (26,  6), 
(39,  4),  (78,  2)}.  The  RHC  and  HH  strategies  were  simulated  1000  times  for  each 
sample  size  m  =  6,  26  or  78  using  XLISP-STAT  (Tierney  1991).  The  estimators 
in  Eq.'s  1,  2,  3,  and  4  were  used  in  computations.  Boxplots  of  the  estimates  of 


total  tree  volume,  Frhc  or  Y^,  and  estimates  of  their  standard  errors, 


v(Xrhc) 


or 


v(Ym) 


are  used  to  summarize  the  simulation  results  (figures  2  and  3). 


Overall,  both  strategies  resulted  in  the  median  of  the  estimates  close  to  the 
parametric  total  r  =  793  m3,  and  the  RHC  estimates  were  slightly  less  variable 

than  the  HH  estimates  (figure  2).  Because  both  Yrhc  and  Ym  are  unbiased 
estimators  with  respect  to  their  sampling  designs,  these  results  were  anticipated. 
As  the  sample  size  increased,  both  sets  of  estimates  became  less  variable 
Variability  of  the  estimates  was  similar  for  m  =  6,  but  the  RHC  strategy  produced 
estimates  with  proportionately  less  variability  as  compared  to  the  HH  strategy  as 
the  sample  size  increased. 

The  estimates  of  the  standard  errors  became  smaller  as  sample  size  increased, 
and  the  estimates  from  the  RHC  strategy  became  proportionately  smaller  than  the 
estimates  from  the  HH  strategy  as  the  sample  size  increased   The  variability  in  the 


437 


o 

CO 

0 

0 

0 

0 

0 

o 

°o                 °  ° 

c- 

M                                       0  o 

°           #  o 
o         0           0  0  °  0 

0°        ft                   °0           o    *  0 
0  « 

0                        o          0     o  0 

°      O                    O     0           0         <>              0  o 

a 

*                        0          „               0  0 

o                                0                o  o 

*>              o  °  o 

o    o          oo              °          0  .  0 

o  °°       (VV        *°  ° 

O 

© 

0 

o 

0 

a 

40                    50                    60                    70                    80  90 

Figure  1. — Geometric  representation  of  the  dispersion  of  "trees."   Each  tree  was 

generated  by  two  independent  random  values  obtained  from  Poisson(60)  +  Uniform(0,  1). 
Circles  illustrate  the  K-circIes  of  the  trees. 


RHC  standard  error  estimates  decreased  proportionately  relative  to  the  HH 
standard  error  estimates  as  sample  size  increased.  Thus  the  standard  error  of  the 
RHC  strategy  (Eq.  2)  tends  to  give  smaller  standard  error  estimates  relative  to  the 
HH  estimates  (Eq.  4),  although  the  practical  difference  appears  minimal  for  the 
generated  Poisson  dispersion  pattern  and  the  normally  distributed  tree  volumes  of 
this  illustration. 


438 


m 
o 
+ 


RHC         HH        RHC         HH        RHC  HH 

///  =  6  m  —  26  //;  =  78 


x 


1  I 


T 


tree  volume 


Figure  2. — Boxplots  of  the  Rao-Hartley-Cochran  (RHC)  and  Hansen- 
Hurwitz  (HH)  estimates  of  total  tree  volume.  The  total  tree  volume  is  793.6 
cubic  meters,  indicated  by  the  horizontal  line.  Sample  size  is  m. 


RHC  HH 

m  =  6 


RHC  HH 

m  =  26 


RHC  HH 

wi  =  78 


□ 

o 


- 


T 


tree  volume 


Figure  3. — Boxplots  of  the  Rao-Hartley-Cochran  (RHC)  and  Hansen- 
Hunvitz  (HH)  standard  error  estimates  of  estimated  total  tree  volume. 
Sample  size  is  m. 


439 


REFERENCES 


Chauduri,  A.  and  Mitra,  J.  1992.  A  note  on  two  variance  estimators  for 
Rao-Hartley-Cochran  estimator.  Communications  in  Statistics  -  Theory  and 
Methods,  21:  3535-3543. 

de  Vries,  P.  G.  1989.  Sampling  theory  for  forest  inventory.  New  York,  NY: 
Springer- Verlag.  399  p. 

Deshpande,  M.  N.  1984.  A  note  on  Rao,  Hartley  and  Cochran's  method.  Sankhya, 
36:114-116. 

Hansen,  M.H.  and  Hurwitz,  W.N.  1943.  On  the  theory  of  sampling  from  finite 
populations.  Annals  of  Mathematical  Statistics,  20:  426-432. 

Ohlsson,  E.  1989.  Variance  estimation  in  the  Rao-Hartley-Cochran  procedure. 
Sankhya,  B,  51:  348-361. 

Rao,  J.N.K.,  Hartley,  H.O.,  and  Cochran,  W.G.  1962.  On  a  simple  procedure  of 
unequal  probability  sampling  without  replacement.  Journal  of  the  Royal 
Statistical  Society  B,  24:  482-491. 

Roesch,  F.  A.,  Jr.,  Green,  E.  J.,  and  Scott,  C.  T.  1993.  An  alternative  view  of 
forest  sampling.  Survey  Methodology,  19:  199-204. 

Singh,  R.  and  Kishmore,  L.  1975.  On  Rao,  Hartley  and  Cochran's  method. 
Sankhya,  C,  37:88-94. 

Tierney,  L.  1990.  LISP-STAT:  An  object-oriented  environment  for  statistical 
computing  and  dynamic  graphics.  New  York,  NY:  John  Wiley.  397  p. 

BIOGRAPHICAL  SKETCH 

Jeffrey  S.  Pontius  is  an  Assistant  Professor  in  the  Department  of  Statistics  and 
the  Center  for  Applied  Statistics,  Kansas  State  University.  His  research  interests 
are  sampling  and  stochastic  processes  in  ecology. 


440 


Increasing  Spatial  Precision  and 
Accuracy  for  Monitoring  Peatlands  in 
Switzerland  by  Remote  Sensing 

Techniques 

Michael  Kohl1 ,  Hans  Jorg  Schnellbacher2 ,  and  Andreas  Griinig3 

ABSTRACT.  -  In  Switzerland  about  1700  peatlands  are  considered  to  be 
'sites  of  national  importance'  and  are  protected  on  a  constitutional  level. 
In  1994  a  peatland  monitoring  program  has  been  initiated.  One  of  the  very 
first  steps  of  this  program  is  to  develop  a  monitoring  design  that  provides 
reliable  and  cost-efficient  information  on  the  current  state  and 
development  of  national  peatlands.  A  combined  photo-/  field  sampling 
design  is  compared  to  a  field-/  based,  one-phase  approach  in  a  Monte 
Carlo  simulation  study.  The  results  show  that  concerning  variance,  bias 
and  root  mean  square  error  the  combined  approach  proved  superior 
compared  to  pure  field  assessments. 


INTRODUCTION 


In  1984  and  1990  respectively,  the  Swiss  national  bog  and  fenland  inventories 
were  carried  out  (Griinig  et  al.,  1986,  Broggi,  1990).  The  area  of  Swiss  mire 
habitats  identified  by  these  two  inventories  totals  approximately  30' 000  ha,  which 
is  less  than  1%  of  the  total  area  of  Switzerland.  These  inventories  included 
approximately  1700  mires  with  a  total  of  20000  ha,  which  are  regarded  to  be  'Sites 
of  National  Importance'  (Griinig,  1994,  Kiittel,  1995).  As  by  federal  legislation 
those  sites  have  to  be  protected  and  monitored,  a  concept  for  monitoring  national 
peatlands  to  meet  the  legal  requirements  had  to  be  developed.  The  focus  of  the 


'  Project  leader.  Inventory  Methods  Research  Group.  Swiss  Federal  Institute  for  Forest.  Snow  and  Landscape  Research. 
8903  Birmensdorf,  Switzerland 

2  Research  forester.  Inventory  Methods  Research  Group.  Swiss  Federal  Institute  for  Forest.  Snow  and  Landscape 
Research.  8903  Birmensdorf,  Switzerland 

3  Project  leader.  Advisory  Service  for  Mire  Conservation.  Swiss  Federal  Institute  for  Forest.  Snow  and  Landscape 
Research.  8903  Birmensdorf.  Switzerland 


441 


monitoring  program  is  to  obtain  results  for  the  entirety  of  Swiss  peatlands  and  to 
describe  the  situation  of  individual  sites  (objects). 

Long  term  monitoring  of  Swiss  peatlands  has  to  cover  two  different  objectives: 
1)  to  detect  changes  in  the  total  area  of  peatlands  and  2)  to  detect  changes  of  the 
area  covered  by  distinct  vegetation  types.  As  vegetation  cover  is  an  indicator  for 
the  quality  of  site  conditions,  vegetation  cover  serves  to  the  purpose  of  monitoring 
site  changes  as  well.  The  methods  applied  to  meet  the  first  objective  have  been 
described  elsewhere  (Kleinn,  1991,  Kohl  and  Kleinn,  1994,  Kohl,  1994)  and  are 
not  subject  of  this  paper. 

Two  monitoring  approaches  will  be  presented:  surveys  based  on  systematically 
distributed  sample  plots,  and  a  two-phase  sampling  design  based  on  the 
combination  of  data  from  aerial  photographs  and  field  assessment.  The  accuracy 
(bias)  of  both  methods  as  well  as  the  precision  for  estimating  current  state  and 
change  is  investigated  by  simulation  studies  and  pilot  surveys  on  a  test  site. 


A  traditional  approach  in  vegetation  surveys  is  to  establish  and  assess  a  set  of 
systematically  distributed  sampling  units  and  remeasure  it  at  successive  occasions. 
In  forest  surveys  this  design  is  known  as  control  sampling  (Schmid,  1967)  or 
continuous  forest  inventory  (Stott,  1947).  On  each  plot  either  the  percentage  of 
area  cover  by  species  or  the  presence  or  absence  of  a  species  can  be  recorded. 
Preliminary  studies  have  shown  that  the  assessment  of  the  percentage  area  cover 
on  sample  plots  is  prone  to  subjective  observer  bias,  which  cannot  be  reduced  by 
sample  size  (see  Gertner  and  Kohl,  1992,  for  details).  Thus  we  will  concentrate  on 
the  second  approach,  which  is  more  robust  in  respect  of  observer  bias. 

The  proportion  of  the  reference  area  covered  by  a  species  can  be  estimated  by 
defining  a  variable  y,  which  can  take  the  value  1  if  the  species  is  observed  on  the 
plot  and  the  value  0  in  all  other  cases.  For  large  populations  the  proportion  P  can 
be  estimated  as  follows: 


SAMPLING  ALTERNATIVES 


[1] 


with  variance 


[2] 


442 


where  n  is  the  number  of  sampling  points  and  yt  is  the  observation  on  plot  i.  The 
variance  of  P  is  estimated  by 


P(l-P) 
n-  1 


If  the  survey  is  based  on  remeasurement  of  the  same  set  of  permanent  plots, 
change  between  two  successive  occasions,  C,  and  its  variance  v(C),  can  be 
calculated  by  taking  the  difference  of  the  proportions  Pi  at  each  occasion. 

^=Y2-fi  =  P2-Pi  ^  t, 

v(t)  =  var(p,)  +  var(p2)  -  2ryx>/var(p1)A/var(p2) 
where 

Yj  =  Pj  =  proportion  at  occasion  i  (i=l,2) 

ryx  =  correlation  coefficient  of  observations  at  occasion  2  and  occasion  1 

In  forest  survey  applications  it  has  often  been  shown  that  cost-efficiency  can  be 
improved  by  combining  information  assessed  in  aerial  photographs  and  field 
assessments.  A  two-phase  (or  double)  sampling  for  stratification  design  has  been 
introduced  in  the  second  Swiss  national  forest  inventory  (Kohl,  1994),  where  in  a 
first  phase  strata  weights  are  assessed  on  a  large  sample  of  aerial  photography 
plots  and  variables  of  interest  are  assessed  on  a  smaller  sample  of  field  plots.  This 
design  is  proposed  as  an  alternative  for  monitoring  the  area  cover  percentage  in  the 
Swiss  peatland  monitoring  program. 

Applying  a  double-sampling  for  stratification  design  the  mean  proportion  of  area 

cover  of  a  plant  species,  Yds  and  its  variance  v(  Yds)  is  estimated  as  follows: 

Yds"  £  7~Yh-P 
h  =  l  n 

v  Yds  "  £  -^T-7v^Yh)+  I  -r—  -r  Yh" Yds 
h=i  n  -  l  n  h=l  n  -1  n 


where 


443 


Yh  =  mean  value  in  stratum  h,  h=l,...L 

v(Yh)  =  variance  of  Yh,  h=l,...L 
■ 

nh  =  number  of  photo  plots  in  stratum  h,  h=l,...L 
n'  =  total  number  of  photo  plots  =  Znh 
L  =  number  of  strata 


DATA 


Among  the  peatlands  covered  by  pilot  surveys  a  mire  called  'Gross  Moos'  in  the 
Schwandital,  Canton  of  Glarus,  has  been  selected  to  provide  basic  data  on  plant 
communities  and  their  spatial  distribution  for  a  simulation  study.  Aerial 
photographs  (CIR,  1:5000)  of  the  Grossmoos  were  taken  in  1994.  The  aerial 
photographs  were  scanned  and  analyzed  by  means  of  digital  photogrammetry  using 
a  HELAVA  DPW  770  station.  The  borderline  of  the  'Gross  Moos'  was  plotted 
and  patches  within  the  total  area  were  assigned  and  delineated  according  to 
differences  in  texture  and  color.  Similarities  of  texture  and  color  were  used  to  form 
nine  strata,  which  can  be  regarded  to  reflect  distinct  plant  communities.  Thus  the 
construction  of  strata  was  not  founded  on  the  occurrence  of  vegetation  types  or 
plant  communities  observed  on  the  ground,  but  exclusively  on  attributes  (texture 
and  color)  assessed  in  aerial  photographs.  Each  of  the  delineated  patches  was 
grouped  into  one  of  the  strata  and  the  results  of  the  photo-interpretation  were 
printed  in  mapped  form.  Based  on  a  preliminary  map  a  vegetation  survey  was 
conducted.  For  the  application  of  a  two-phase  design,  however,  only  strata  have  to 
be  assigned  to  photo  plots,  but  no  map  is  required.  For  each  patch  a  field 
assessment  was  conducted  and  the  plant  species  observed  and  their  cover 
recorded.  The  cover  of  each  plant  species  within  each  patch,  p,  was  assessed  in 
five  classes: 

-  not  present  -  0%  <  p  <  0. 1  % 

-  0.1%  <  p  <  1%  -  1%  <  p  <  10% 

-  >  10% 

The  results  of  the  vegetation  survey  were  lateron  used  to  assign  vegetation  types 
to  the  strata  formed  by  aerial  photo-interpretation.  For  the  simulation  study  six 
plant  species  reflecting  different  spatial  distributions  and  abundance  classes  were 
selected.  Table  1  summarizes  the  spatial  distribution  and  the  proportion  of  area 
cover  for  the  selected  species. 


444 


Table  1.  -  Selected  species  and  corresponding  spatial  distribution  and  area  cover  of 
patches,  in  which  species  have  been  observed.  


Species 


spatial  distribution  of  patches  proportion  of  area  cover 
 within  patches  


Andromeda  polifolia 

Anthoxantum  odoratum 
Scheuchzeria  palustris 

Senecio  alpinus 
Tofieldia  calyculata 

Vaccinium  myrtillus 


present  on  contagious  part  of  the  all  classes 
area 

clustered  frequent,  majority  >0.1% 

clustered  in  small  parts  of  the  frequent,  majority  >0.1% 
area 

clustered  at  the  margin  rare,  majority  <  1% 

present  on  small  contagious  part  frequent,  majority  >0. 1  % 
of  the  area 

present  on  majority  of  patches  frequent,  majority  >0. 1  % 


SIMULATION  STUDY 


In  a  Monte  Carlo  simulation  study  1 -phase  and  2-phase  sampling  designs  were 
tested.  The  field  plots  as  well  as  the  photo  plots  were  systematically  distributed 
over  the  entire  area.  In  two  phase  designs  the  photo  plots  were  used  to  estimate 
the  stratum  weights.  On  field  plots  the  presence  or  absence  of  each  of  the  six 
species  was  recorded.  Table  2  summarizes  the  sampling  designs  and  grid  densities 
included  in  the  simulation  study. 


Table  2.  -  Design  alternatives  applied  in  the  simulation  study 

Sampling  design                            Grid  density  Sai 

nple  size 

1 -phase  sampling                                               5*  5m 

6369 

10*  10  m 

1599 

20  *  20  m 

402 

2-phase  sampling  for  stratification              1st  phase3:  5  *  5  m 

1st  phase3: 

6369 

2ndphaseb:  10  *  10  m 

2nd  phase6: 

1599 

lst-phasea:  10  *  10  m 

1st  phase3: 

1599 

2nd  phase5:  20  *  20  m 

2nd  phase1 

':  402 

3  aerial  photography  b  field  assessment 


The  simulation  study  was  based  on  the  map  obtained  by  the  aerial  photo- 
interpretation  and  the  vegetation  survey.  In  the  simulation  study  an  exact 
proportion  of  area  cover  within  the  observed  cover  class  was  randomly  selected 
and  assigned  to  each  patch.  To  simulate  the  spatial  distribution  within  each  patch  a 
matrix  of  pixels  with  2. 5*2. 5m  spatial  resolution  was  laid  over  the  entire  area  of 
the  'Gross  Moos'.  For  each  pixel  a  random  number  was  drawn  and  compared  to 
the  proportion  of  area  coverage  of  a  species  within  the  patch,  in  which  the  pixel  is 
located.  If  the  random  number  was  smaller  than  the  randomly  assigned  proportion, 
the  species  under  concern  was  assumed  to  be  present  on  the  plot,  in  all  other  cases 
the  species  was  assumed  not  to  be  present  on  the  pixel.  To  test  the  efficiency  of 


445 


the  5  design  alternatives  for  the  estimation  of  change,  the  area  cover  of  each 
species  was  altered  by  10  percent.  Thus  the  absolute  value  of  changes  of  percent 
area  cover  was  higher  for  frequent  species  than  for  species  with  low  percent  area 
cover.  Based  on  the  randomly  generated  maps  a  second  map  was  produced,  in 
which  those  pixels  were  randomly  selected,  where  the  presence  or  absence  of 
species  changed.  This  procedure  was  repeated  for  each  species.  The  original 
stratification  with  3  strata  was  retained  for  all  maps. 

New  maps  for  current  state  and  changes  were  generated  for  each  of  the  1000 
iterations  and  the  5  sampling  alternatives  were  simulated  in  each  map.  For  each 
sampling  alternative,  each  species,  and  each  iteration  the  following  parameters 
were  estimated: 

-  current  state  of  percent  area  cover  and  its  variance  using  the  equations  presented 
above, 

-  the  bias,  where  bias  =  |  Ptme  -  P  I ,  and 


According  to  Cochran  (1977)  the  RMSE  can  be  used  as  a  criterion  of  the 
accuracy  of  an  estimator. 


In  Table  3  to  Table  8  the  results  for  the  six  species  are  presented.  For  each  of  the 
5  design  alternatives  the  estimates  proportion,  P,  the  bias,  the  standard  error  of  P, 
s(P),  and  the  root  mean  square  error,  rmse,  are  given. 

The  standard  error  of  (P)  is  much  smaller  for  the  2-phase  sampling  designs  than 
for  the  1 -phase  designs.  This  holds  for  both,  current  values  and  change,  and  can  be 
observed  independent  of  the  spatial  distribution  and  the  frequency  of  the  species. 
Even  for  the  denser  grid  size  the  variance  of  the  1 -phase  designs  is  larger  than  for 
the  coarsed  grids  of  the  2-phase  design.  For  the  estimation  of  current  state  and 
change  the  bias  is  smaller  for  the  2-phase  designs  with  one  exception  (10*10/ 
5* 5m  grid  for  vaccinium  myrtillus).  The  size  of  the  bias  of  the  1 -phase  and  2-phase 
designs  is  similar,  when  the  terrestrial  grid  sizes  of  the  1 -phase  designs  and  the  grid 
size  of  the  photo  phase  in  2-phase  designs  are  equal.  The  same  effect  can  be 
observed  for  the  root  mean  square  error.  The  studied  2-phase  designs  arrive  at  the 
same  root  mean  square  with  one  quarter  of  the  field  plots  of  a  1 -phase  design.  As 
the  ratio  of  change  was  the  same  in  all  strata,  an  improvement  is  likely  for 
situations,  where  the  changes  of  cover  vary  in  individual  strata. 


the  root  mean  square  error  RMSE,  where  RMSE  =  y  v(P)  +  bias 


RESULTS 


446 


Table  3:  Scheuchzeria  palustris  (Ptnie  =  1.1416,  Ctrue  =  0.1041) 


Grid  size  [m] 

field: 

5*5 

10*10 

20*20 

10*10 

20*20 

photo: 

5*5 

10*10 

current  state 

P 

1.2199 

1.0175 

1.0445 

1. 

0349 

1.1521 

bias 

0.0782 

0. 1241 

0.0971 

0. 

1067 

0.0105 

s(P) 

0.1371 

0.2493 

0.4970 

0.0015 

0.0023 

rmse 

0.1711 

0.3157 

0.6078 

0. 

1564 

0.3104 

changes 

C 

0.1114 

0.0939 

0.0970 

0.0956 

0.1067 

bias 

0.0073 

0.0101 

0.0070 

0.0085 

0.0027 

s(C) 

0.0409 

0.0649 

0.0883 

0.0001 

0.0002 

rmse 

0.0525 

0.1003 

0.1768 

0.0616 

0.1389 

Table  4:  Vaccinium  myrtillus  (Ptrue 

=  8.3311,  Ctrue  =  0.8335) 

Grid  size  [m] 

Field: 

5*5 

10*10 

20*20 

10*10 

20*20 

Photo: 

5*5 

10*10 

Current  state 

P 

8.3827 

8.2727 

8.1731 

8.1376 

8.1916 

bias 

0.0516 

0.0584 

0.1580 

0.1935 

0.1395 

s(P) 

0.3466 

0.6875 

1.3625 

0.0173 

0.0234 

rmse 

0.3972 

0.8147 

1.6503 

0.3973 

0.7934 

Change 

C 

0.8393 

0.8291 

0.8386 

0.8153 

0.8386 

bias 

0.0057 

0.0045 

0.0050 

0.0183 

0.0050 

s(C) 

0.1139 

0.2243 

0.4352 

0.0022 

0.0021 

rmse 

0.1442 

0.2993 

0.5890 

0.1707 

0.3398 

Table  5:  Andromeda  polifolia  (Ptrut 

=  0.5267,  Ctrue  =  0.0501) 

Grid  size  [m] 

Field: 

5*5 

10*10 

20*20 

10*10 

20*20 

Photo: 

5*5 

10*10 

Current  state 

P 

0.5353 

0.5353 

0.7007 

0.5348 

0.6274 

bias 

0.0086 

0.0085 

0.1740 

0.0081 

0.1007 

s(P) 

0.0908 

0.1795 

0.3975 

0.0045 

0.0012 

rmse 

0.1077 

0.2237 

0.5165 

0.1145 

0.2494 

Change 

A 

C 

0.0496 

0.0470 

0.0622 

0.0469 

0.0564 

bias 

0.0005 

0.0031 

0.0121 

0.0031 

0.0063 

s(t) 

0.0263 

0.0382 

0.0577 

0.0002 

0.0001 

rmse 

0.0343 

0.0671 

0.1147 

0.0425 

0.0821 

447 


Table  6:  Tofieldia  calyculata  (Ptrue  =  0.0962,  Ctnie  =  0.0094) 


Grid  size  [m] 


Field: 

5*5 

10*10 

20*20 

10*10 

20*20 

Photo: 

5*5 

10*10 

V^UllClll  MalC 

TJ 

r 

0  0Q54 

0  10Q4 

0  1418 

U.I'tlO 

0  107^ 

O  1489 

bias 

0.0008 

0.0133 

0.0456 

0.0112 

0.0521 

s(P) 

0.0378 

0.0720 

0.1214 

0.0007 

0.0012 

rmse 

0.0491 

0.1059 

0.2116 

0.0623 

0.1583 

Change 

C 

0.0097 

0.0124 

0.0172 

0.0122 

0.0180 

bias 

0.0003 

0.0031 

0.0078 

0.0028 

0.0087 

sit) 

0.0082 

0.0117 

0.0170 

0.0001 

0.0002 

rmse 

0.0134 

0.0227 

0.0320 

0.0170 

0.0256 

Table  7:  Anthoxantum  odoratum  (Ptnie  =  3.0541, 

Ctrue  =  -0.3047) 

Grid  size  [m] 

Field: 

5*5" 

10*10 

20*20 

10*10 



20*20 

Photo: 

5*5 

10*10 

Current  state 

P 

3.0558 

3.0882 

3.7194 

3.0964 

3  7019 

bias 

0.0017 

0.0341 

0.6653 

0.0423 

0.6479 

s(P) 

0.2151 

0.4312 

0.9394 

0.0038 

0.0081 

rmse 

0.2470 

0.5096 

1.2524 

0.2313 

0.7371 

Change 

C 

0.3082 

0.3091 

0.3776 

0.3098 

0.3756 

bias 

0.0035 

0.0044 

0.0729 

0.0051 

0.0709 

s(C) 

0.0689 

0.1352 

0.2646 

0.0004 

0.0007 

rmse 

0.0878 

0.1779 

0.3856 

0.0986 

0.2282 

Table  8:  Senecio  alpinus  (Ptrue  =  1.9300,  Ctrue  =  0.1767) 

Grid  size  [ml 

Field: 

5*5 

10*10 

20*20 

10*10 

20*20 

Photo: 

5*5 

10*10 

Current  state 

P 

1.9203 

1.9548 

2.1542 

1.9571 

2.2573 

bias 

0.0097 

0.0248 

0.2242 

0.0271 

0.3273 

s(P) 

0.1717 

0.3451 

0.7191 

0.0019 

0.0045 

rmse 

0.1947 

0.4011 

0.8690 

0.1730 

0.4832 

Change 

C 

0.1767 

0.1828 

0.2119 

0.1830 

0.2227 

bias 

0.0000 

0.0061 

0.0352 

0.0063 

0.0459 

s(t) 

0.0520 

0.1009 

0.1713 

0.0002 

0.0004 

rmse 

0.0661 

0.1374 

0.2818 

0.0788 

0.1891 

CONCLUSION 


The  results  of  the  simulation  study  showed  that  the  same  root  mean  square  error 
can  be  achieved  by  2-phase  sampling  designs,  which  use  a  large  set  of  photo  plots 
to  estimate  strata  sizes  and  a  subsample  of  field  plots,  with  about  lA  of  the  field 


448 


plots  of  a  comparable  1 -phase  design.  The  efficiency  of  a  2-phase  design  proofed 
to  be  superior. 

However,  the  optimal  design  can  only  be  selected,  when  cost-efficiency  is 
concerned.  The  optimal  design  among  the  design  alternative  is  the  one,  which 
either  leads  to  the  smallest  variance  for  given  cost  or  which  results  in  the  smallest 
cost  for  a  desired  variance.  For  a  desired  variance  the  total  cost  of  obtaining  aerial 
photographs,  aerial  photo  interpretation  and  amortization  of  photo  interpretation 
equipment  should  be  less  than  75  percent  of  the  total  cost  of  a  strictly  field  based 
assessment,  referring  to  the  results  obtained  in  the  simulation  study.  A  2-phase 
design  would  be  superior  in  terms  of  cost-efficiency,  if  the  cost-ratio  of  field  and 
photo  plots  would  be  1.33:1.  Experience  has  shown  that  the  cost  ratio  is  much 
better  referring  to  photo  samples.  Taking  into  account  other  side  effects  of  aerial 
photography  such  as  the  documentation  of  a  historical  situation,  providing  a  piece 
of  evidence  for  later  investigations,  and  allowing  retrospective  analysis  at  later 
points  in  time,  the  superiority  of  the  2-phase  approach  is  evident. 

In  many  situations  vegetation  surveys  are  purely  based  on  field  assessments.  The 
data  recorded  for  a  single  plot  are  extrapolated  to  the  area  represented  by  the  plot, 
i.e.  the  surrounding  area,  and  used  to  produce  pixel  maps.  The  grid  density  of  1- 
phase  sampling  designs  has  to  be  fairly  large  to  give  a  realistic  situation  of  the  spa- 
tial distribution  and  frequency  of  species  in  the  area  under  concern.  Especially,  the 
necessary  sample  size  to  detect  sensitive  (and  meaningful)  changes  of  spatial  pat- 
ters should  not  be  underestimated.  A  stratification  in  aerial  photographs  provides  a 
more  realistic  picture  of  the  spatial  distribution  of  vegetation  types.  If  a  2-phase 
design  is  applied,  cost  efficiency  for  the  estimation  of  current  state  and  changes  is 
better  compared  to  1 -phase  designs.  Thus,  whenever  possible,  the  combination  of 
field  assessments  and  photo  interpretation  should  be  given  preference. 

ACKNOWLEDGMENT 


We  would  like  to  thank  Dr.  Bernhard  Oester,  WSL,  for  providing  the  results  of 
the  photo-interpretation.  Dr.  Risto  Paivinen,  European  Forest  Institute  (EFI),  and 
Dr.  Peter  Brassel,  WSL,  reviewed  the  manuscript  and  gave  helpful  comments. 

REFERENCES 

Broggi,  MF.  (Ed),  1990:  Inventar  der  Flachmoore  von  nationaler  Bedeutung.  Entwurf  fur  die 
Vemehmlassung.  Bern,  Bundesarrt  fur  Umwek,  Wald  und  Landschaft  (BUWAL)  79  p. 

Cochran,  W.G.  1977.  Sampling  techniques,  J.  Wiley  &  Sons,  New  York.  428  p. 

Gertner,  G.Z.,  Kohl,  M.  1992.  An  assessment  of  some  nonsampling  errors  in  a 
national  survey  using  an  error  budget.  Forest  Science.  Vol.  38,  No.  3.  525-538 


449 


Griinig,  A;  Vetterli,L;Wi]di,C).,  1986:  Die  Hoch-  und  Uebergangsmoore  der  Schweiz  - 
eine  Inventarauswertung.  Ber.  Eidgenbss.  Forschanst  Wald  Schnee  Lardsch  281, 62  p. 

Griinig,  A.  (Ed.),  1994:  Mires  and  Man.  Mire  conservation  in  a  densely  populated 
country  -  the  Swiss  experience.  Excursion  guide  and  symposium  proceedings 
of  the  5th  field  symposium  of  the  International  Mire  Conservation  Group 
(IMCG)  to  Switzerland  1992.  Birmensdorf,  Swiss  Federal  Institute  for  Forest, 
Snow  and  Landscape  Research.  415  p. 

Kish,  L.  1965.  Survey  sampling,  J.  Wiley  &  Sons,  New  York,  643  p. 

Kleinn,  C.  1991.  Der  Fehler  von  Flachenschatzungen  mit  Punktrastern  und  linien- 
formigen  Stichproben,  Universitat  Freiburg.,  Mitt.  Abt.  Forstl.  Biom.,  Nr.  91-1 

Kohl,  M.  1994.  Statistisches  Design  fur  das  zweite  Schweizerische  Landesforst- 
inventar:  Ein  Folgeinventurkonzept  unter  Verwendung  von  Luftbildern  und 
terrestrischen  Aufnahmen,  Mitteilungen  der  Eidgenossischen  Forschungsanstalt 
fur  Wald,  Schnee  und  Landschaft,  Birmensdorf,  Band  69,  Heft  1,  141  S. 

Kohl,  M.,  Kleinn,  C.  1994.  Stichprobeninventuren  zur  Waldflachenschatzung. 
Allgemeine  Forst-  und  Jagdzeitung,  165  (12):  229-231 

Kttttel,  M.,  1995:  Moorschutz  in  der  Schweiz  -  Stand  und  Ziele.Telma25. 177-192. 

Schmid,  P.  1967.  Die  Weiterentwicklung  der  Leistungskontrolle  in  der  Schweiz. 
Wiss.Z.Tech.  Univ.  Dresden,  Jg.  16/2 

Stott,  CB.  1947.  Permanent  Growth  and  Monitoring  Plots  in  Half  the  Time.  J.for.  37: 669-673 


BIOGRAPHICAL  SKETCH 


Michael  Kohl  is  project  leader  and  head  of  the  Inventory  Methods  Research 
Group,  Swiss  Federal  Institute  for  Forest,  Snow  and  Landscape  Research  (WSL). 
He  graduated  from  University  of  Freiburg,  Germany,  in  1983  and  holds  a  diploma 
in  forestry  and  a  PhD  in  biometry  (1986).  He  is  lecturer  at  the  Swiss  Federal 
Institute  of  Technology  (ETH),  Zurich,  where  he  habilitated  in  1993.  Michael  is 
responsible  for  the  statistical  and  remote  sensing  methods  of  the  Swiss  National 
Forest  Inventory  and  is  leader  of  IUFRO  S4.ll  (Statistics,  Mathematics  and 
Computers). 

Hans  Jorg  Schnellbacher  is  a  research  forester  with  the  Inventory  Methods 
Research  Group,  WSL.  He  graduated  from  University  of  Freiburg,  Germany,  in 
1992  and  holds  a  diploma  in  forestry.  Hans  Jorg  provides  statistical  consulting  for 
the  Swiss  National  Forest  Inventory  and  is  responsible  for  the  data-base 
management  and  the  statistical  analysis  of  the  inventory  data. 

Andreas  Griinig  is  project  leader  and  head  of  the  Advisory  Service  for  Mire 
Conservation,  WSL.  He  graduated  at  the  Swiss  Federal  Institute  for  Technology 
(ETH),  Zurich  and  holds  a  diploma  as  a  scientist  (biologist).  Andreas  is  responsible 
for  the  Swiss  Mire  Monitoring  Programme. 


450 


Spatial  and  Temporal  Models  in 
Contextual  Classification 


Bo  Ranneby 


Abstract.  Different  kinds  of  spatial  and  temporal  models  and  their  use  for 
contextual  classification  of  satellite  images  are  discussed.  Models 
describing  the  ground  cover  and  its  changes  over  time  and  models  of  the 
spectral  signatures  in  multitemporal  images  are  studied.  Isotropic 
autocorrelation  models  as  well  as  mosaic  models  are  used  for 
quantification  of  the  variation  in  space  and  time. 

For  the  different  models  parameter  estimation,  robustness  properties,  and 
classification  performance  are  examined.  The  use  of  appropriate 
estimation  techniques  is  crucial  for  the  result  of  the  classification.  The  use 
of  multitemporal  images  makes  it  easier  to  construct  robust  estimation 
methods,  and  also  improves  the  classification  results.  

INTRODUCTION 

The  purpose  of  this  paper  is  to  discuss  some  models  and  estimation  problems  which 
appear  in  the  classification  of  remote  sensing  images.  The  true  image  on  a  two- 
dimensional  regular  lattice  D  will  be  denoted  by  {C(s),s  e  D}.  The  measurement  of 

pixel  s  is  given  by  Z(s).  Unfortunately,  there  is  no  one  to  one  correspondence  between 
the  C-  and  Z-process,  but  they  are  contaminated  with  noise.  The  classification 
algorithms  referred  to  in  the  present  paper  relies  on  Bayesian  methods.  Let  7i(c) 
denote  the  prior  distribution  of  the  C-process  and  let  f(z  I C)  denote  the  conditional 
distribution  of  Z  given  C.  Then  the  posterior  distribution  of  C  given  Z  is 
p(C  I Z)  oc  71(C)  f(Z  I C).  The  maximum  a  posteriori  estimator  is  the  one  that  maximizes 
p(C  I Z).  The  crucial  steps  in  the  derivation  of  the  classification  algorithm  are 
modelling  the  different  processes. 

MOSAIC-MODELS 

For  many  of  the  applications  we  have  in  mind  mosaic-models  or  models 
related  to  mosaic  models  will  give  a  realistic  description  of  true  classes  of  the 
image  and  the  development  over  time  of  these  classes.  As  a  starting  point  to  the 


Department  of  Forest  Resource  Management  and  Geomatics 
Swedish  University  of  Agricultural  Sciences 
S-901  83  UMEA 
Sweden 


451 


study  of  mosaic-models  we  consider  stochastic  methods  of  dividing  the  plane 
into  contiguous  subsets  called  cells. 

A  simple  (to  handle  mathematically)  mosaic  is  the  following  line-mosaic,  see 
Pielou  (1964)  and  Switzer  (1965).  The  plane  is  partitoned  into  convex  cells  by 
locating  straight  lines  at  random  according  to  the  following  procedure.  Choose 
points  (a,9),  in  the  infinite  strip  0  <  9  <  n,  according  to  a  Poissonprocess  with 
intensity  y  points  per  area  unit  and  associate  with  each  point  (a,9)  the  line 

y  cos  9  +  xsin9-a  =  0 

Each  cell  is  then  independently  assigned  its  color  according  to  a  probability  law, 
which  selects  class  i  with  probability  p;.  If  there  is  only  two  colors  to  choose 
between  we  get  the  Pielou-Switzer  model. 

This  model  has  also  been  used  by  Owen  (1984)  in  contextual  classification.  If 
the  intensity  parameter  of  the  Poisson  field  is  small  relative  to  the  size  of  the 
pixels  then: 

(i)  only  three  types  of  patterns  X,L,  and  T  are  possible  in  the  first  order 
neighborhoods  (s,  Ns,  Ss,  Es,Ws). 


All  rotations  are  assumed  to  be  equally  likely. 

(ii)  each  neighborhood  has  the  same  probability  of  intersecting  one  boundary 
line,  and  small  probability  of  intersecting  two  or  more  boundaries. 

(iii)  the  portion  of  a  boundary  within  a  neighborhood  is  a  line  segment,  so  that  a 
single  boundary  can  induce  only  an  L-  or  a  T-pattern. 

To  this  model  the  variation  in  time  is  naturally  incorporated  by  independently 
assigning  a  stochastic  process  ^(t)  to  each  cell.  The  processes  give  the  color  of 
the  cells  at  different  times. 

As  the  size  of  forest  stands  is  large  compared  to  the  pixel  size  this  kind  of  models 
will  be  useful  in  contextual  classification  of  forested  areas. 

Let  us  now  consider  another  way  of  generating  the  cell  partition.  Consider 
points  (centers)  located  in  the  plane  according  to  a  Poissonprocess  with  density 
X.  Attach  to  each  cell  its  Dirichlet  cell  consisting  of  those  points  that  are  closer 
to  this  center  than  to  any  other.  The  Dirichlet  cells  of  a  realization  of  a  Poisson 
process  constitute  a  random  division  of  the  plane  into  polygons  varying  in  size 
and  shape.  The  variance  of  the  area  of  a  cell  is  0.2801 8/A2,  see  Matern  (1979). 
The  mosaic,  and  its  variation  over  time,  is  then  obtained  by  independently 
assigning  a  stochastic  process  to  each  cell  giving  the  color  of  the  cell  at  different 
times. 


X  pattern 


T  pattern 

j 


L  pattern 

j 


452 


For  space-time  models  of  this  kind  it  holds  that  the  correlation  in  space  and 
time  is  the  product  of  the  spatial  correlation  and  the  correlation  in  time  given  by 
the  correlation  of  the  process  ^(t).  As  shown  in  Ranneby  (1982)  this  holds  for  a 
rather  large  class  of  models.  If  we  have  a  partition  into  cells  so  the  cell-model  is 
isotropic  and  we  to  each  cell,  independently  of  the  cell  partition  and  of  other 
cells,  have  attached  stochastic  processes  4iW,  ^Wv  to  the  cells,  then  the 
correlation  function  of  the  variation  in  space  and  time  is  given  by  the  product  of 
the  correlation  function  of  the  variation  in  space  and  the  correlation  function  of 
the  variation  in  time.  It  is  crucial  that  the  ^-processes  have  the  same  distribution. 
The  result  can  be  extended  to  models  where  neighboring  cells  may  be  dependent. 

The  mosaic  models  above  describe  the  variation  on  the  ground.  An  extremely 
simple  model  describing  probabilities  of  different  class  configuration  are 
obtained  by  assuming  conditional  independence.  The  conditional  probability  of 
getting  class  configurations  a,b,c,d  given  class  k  in  the  centre  pixel  is  given  by: 

g(a,b,c.d  I  k)  =  p(a  I  k)  p(bl  k)  p  (c  I  k)  p  (d  I  k) 

This  model  has  been  used  by  Haslett  (1985)  for  contextual  classification.  It  is 
easily  extended  to  images  from  several  occasions,  see  Flygare  (1993). 

MODELS  FOR  SPECTRAL  REFLECTANCES 

For  modelling  the  spectral  reflectances  autocorrelation  models  have  shown  to 
be  useful.  In  most  cases  these  models  are  assumed  to  be  multivariate  Gaussian 
processes.  The  vector  X(s)  of  observed  reflectances  for  pixel  s  at  time  t  is 
assumed  to  be  a  sum  of  two  independent  variables  Y(s)  and  s(s).  The  s-process 
is  autocorrelated  noise,  while  the  Y-process  gives  the  true  signal.  Usually  we 
have  images  from  one  or  two  occacions.  When  we  have  two  images  t  equals  1  or 
2.  When  we  have  only  one  image  the  index  t  is  omitted.  The  following 
assumptions  will  be  made  for  the  autocorrelation  models. 

A, :      The  random  fields  {Y(s),s  e  D}  and  {s(s),  s  e  D}  are  independent 

A2:      The  random  field  {s(s),  seD)  is  Gaussian  with  expectation  zero  and 

Cov(s(s),  s  (s)  =  p  9Z,  where  d  =  |  s-Sj  I  is  the  distance  between  s  and  S\. 

A3:      The  random  field  {Y(s),s  e  D}  is  Gaussian  with  expectation  zero  and 
co variance  matrix  (1  -  9)X  and  Cov(Y(s),  Y  (s,)  =  0,         s  *  s, 

The  parameter  9,  0  <  9  <  1,  is  a  proportion  parameter  telling  us  the  amount  of 
variation  in  X(s)  coming  from  e(s)  vs  Y(s)  and  p.  0  <  p  <  1,  is  the  spatial 


453 


autocorrelation  between  noise  vectors  e(s)  and  s^)  for  pixels  a  unit  distance 
apart. 

The  assumptions  ArA3  are  enough  for  one  image  models.  For  two  image 
models  the  following  additional  assumptions  are  made: 

A4:      The  three  random  fields  {(Y(s,2),  Y(s,l)),s  e  Z)},  {e(s,2\s  <e  D) 
and  {e(s,l),s  €D}  are  independent. 

A5:      The  random  fields  {s(s,2),s  e  D}  and  {s(s,l),s  e  D}  are  Gaussian 
with  expectation  zero  and  with  the  same  covariance  matrix  6X- 

A6:      (Y(s,2),  Y(s,l))  and  (Y(s1?2),  Y(sbl))  are  independent  for  s^Sj 

A7:      (Y(s,2),  Y(s,l))  have  Gaussian  distributions  with  zero  expectation  and 
covariance  matrix  (1  -  9)Z  for  (Y(s,2)  and  Y(s,l)  and  with  cross 

T 

covariance  function  Cov(Y(s,2),  Y\s,l)  =  x(l  -  9)E- 
To  make  these  models  operational  the  different  parameters  have  to  be  estimated. 

ESTIMATION  PROBLEMS 

There  are  interesting  estimation  problems  associated  with  several  of  the  models 
described  above.  Usually  the  training  sets  for  the  different  classes  are  selected  by 
subjective  methods  -  areas  which  are  typical  for  the  different  classes  are  chosen.  This 
implies  that  the  means  may  be  unbiasedly  estimated  but  the  variation  within  the 
classes  may  be  heavily  underestimated.  If  so,  too  many  pixels  will  not  be  "typical"  for 
any  class,  which  may  result  in  an  unnecessary  large  "doubt"-class.  If  the  training  sets 
are  chosen  by  objective  methods  this  problem  is  avoided. 

For  good  performance  of  classification  methods  based  on  the  autocorrelation  model 
it  is  necessary  to  have  almost  unbiased  estimates  of  the  parameters.  If  the  noise- 
process  had  been  observable  the  autocorrelation  parameters  had  been  estimated  by 
replacing  the  theoretical  moments  by  their  empirical  correspondence.  Since  we  can 
only  observe  the  X-process  there  is  a  risk  that  the  autocorrelations  in  the  noise-process 
are  contaminated  with  the  autocorrelations  for  the  true  spectral  reflectances.  To  avoid 
this  homogeneous  pixelcrosses  (the  pixel  and  its  N-,  S-,  W-,  and  E-neighbors)  should 
be  used.  If  the  spectral  reflectances  are  the  same  for  all  pixels,  then  as  we  want,  it  is 
only  the  noise-process,  which  causes  the  autocorrelations.  This  is  also  why  Hjort  et  al 
(1984)  recommends  homogeneous  pixel  crosses.  This  is  the  best  we  can  do  if  only  one 
image  is  available. 

If  images  from  two  occasions,  close  in  time,  are  available  it  is  possible  to  base  our 
estimates  on  the  difference  of  the  two  images.  Taking  the  differences  over  areas  with 
slow  changes  we  get  under  reasonable  assumptions,  the  difference  between  the 


454 


n(d)(i-e)X+PdeZ- 


autocorrelated  noise  for  the  two  images.  This  has  been  studied  in  Flygare  (1996), 
where  she  discusses  some  robustness  properties  and  shows  strong  consistency  of  the 
autocorrelation  estimates. 

Let  us  have  a  closer  look  at  the  robustness  properties  for  the  two  methods.  Suppose 
that  assumption  A3  in  the  previous  section  is  not  satisfied  but  that  Cov(Y(s),  Y(s,))  = 
r)(d)(l  -  6)Z,  where  r)(d)  is  an  unknown  positive  correlation  function  with  n(0)  =  1 .  If 
A3  holds  then  r)(d)  =  0  for  s  *  s,. 

For  the  one  image  method  we  get  that 

—  X(Xk  (s)  -  u(k))(Xk  (S, )  -  u<k))T 

As  r|(d)  in  most  applications  is  larger  than  zero  this  results  in  an  overestimation. 
For  the  difference  method  the  situation  is  somewhat  different.  Here  we  get  the 
following  expected  value  for  our  estimate: 

2(1  -x)r|(d)(l  -6)I  +  2pdei. 

If  i  is  close  to  1  we  see  that  the  expected  value  is  approximately  2pd0Z,  which  is  what 
we  get  when  all  assumptions  are  fulfilled.  As  a  consequence  the  difference  method  is 
less  sensitive  to  errors  in  the  model. 

MIXTURES  OF  MULTINORMAL  DISTRIBUTIONS 

In  Taxt  et  al  (1991)  mixtures  of  normal  distributions  are  used  for  classification  of 
images.  They  have  rather  frequently  observed  bimodal  and  multimodal  empirical 
distributions  for  feature  vectors  from  the  same  class,  when  working  with  remote 
sensing  images.  If  the  training  sets  are  objectively  selected  there  are  reasons  to  believe 
that  empirical  distributions  of  this  kind  will  be  more  common.  Taxt  et  al  (1991)  used 
linear  combinations  of  two  multinomial  distributions  to  model  the  distributions  for  the 
different  classes.  A  typical  model  for  the  density  function  f(x)  of  the  feature  vector  is 
given  by 

f(x)  =  (1  -  7i)Nd  (u„K)(x)  +  7iNd  (u2,M)(x) 

The  estimation  of  the  parameters  in  such  distributions  is  known  to  be  a  difficult 
problem,  both  from  statistical  and  algorithmic  points  of  view. 

Example.  Let  ^,^2—  be  i.i.d.  observations  from  a  mixture  of  two  bivariate  normal 
distributions.  The  density  function  f(x,y,0)  is  given  by 

f(x.y,9)  =  p  Hx,y,\LXi\L2,a\&2,P\)  +  0-p)  h(x,y,  \x3,\xA,G3,c4,p2), 


455 


where  h  is  the  density  function  for  a  bivariate  normal  distribution  with  parameters 
indicated  by  the  notation.  Say  that  we  have  observations  (xby  i),(x2,y2)...,  (xn,y  n).  If 
we  put  m  =  xls  ja2  =  yi  and  let  ab(or  a2)  go  to  zero,  then  the  likelihood  function  tends 
to  infinity.  Consequently,  the  ML-method  is  not  suitable. 

The  same  problem  appears  also  for  univariate  distributions.  In  Ranneby  (1984)  a 
new  estimation  method,  the  Maximum  Spacing  (MSP)  Method,  was  introduced  which 
gives  consistent  estimators  also  for  distributions  of  this  kind.  The  method  is  closely 
related  to  the  Maximum  Likelihood  Method  -  both  can  be  derived  as  approximations 
of  the  Kullback-Leibler  information.  The  MSP-method  is  constructed  so  that  each 
contribution  is  bounded  from  above  for  any  parameter  combination. 

The  same  estimation  method  has  been  used  by  Cheng  &  Amin  (1983)  for  three- 
parameter  Weibull  and  gamma  models.  The  method  is  also  discussed  in  Titterington 
(1985),  where  he  states  that  "in  principle,  of  course  it  would  be  possible  to  treat 
multivariate  data  by  grouping:  although  definition  of  the  multinomial  cells  would  be 
more  awkward." 

Since  we  do  not  have  any  natural  order  relation  in  Rd  when  d  >  1 ,  it  is  not  at  all 
obvious  how  the  generalization  shall  be  carried  out.  In  the  univariate  case  we  attach  to 
each  observation  ^  the  variable  r\{  (n)  =  (n  +  1)  "distance  to  the  nearest  observation  to 
the  right  of 

In  Rd,  d  >  1,  we  cannot  use  r)j(n).  However,  to  each  observation,  ^  we  can  attach 
its  Dirichlet  cell.  The  Dirichlet  cell  Dn(^)  surrounding  §j  consist  of  all  points  x  e 
Rd  which  are  closer  to  §i  than  to  any  other  observation. 

Let  5i>^2—  De  i-i-<i-  random  vectors  in  Rd  with  true  density  function  g(x)  and  suppose 

we  assign  a  model  with  density  functions  {f(x,9);6  e0},  where  0cRq.  Let  A  be 
a  measurable  subset  of  Rd.  Then  Pq  (A)  denotes  the  probability  that  ^  belongs  to  the 
set  A  calculated  under  the  asumption  that  the  density  function  is  given  by  f(x,  0)  (i.e. 

Pq  (A)  =  j*f (x,  0)dx).  The  true  probablity  distribution  is  denoted  by  P0  (•)• 

A 

The  Dirichlet  cells  Dn(£j),  i  =  l,2,...,n  split  Rd  into  n  identically  distributed 

n 

random  sets.  Since  ^P0(Dn(^j))  =  1 ,  symmetry  arguments  give 

i=l 

E[P0(Dn(5i))l  =  -  or  P0(Dn  (^  ,))  =  -.  Since  it  also  holds  that 
J    n  n 

P0(Dn^i))  =  |Dn(5.)|g^i), 

we  get  that  g(^)  =  1/n  I  Dn(^)  I .  Here  I  Dnfe)  I  denotes  the  "volume  of  the  set  Dn(^)". 


456 


Consequently,  the  random  variable  n|  Dn(£i)l  may  be  interpreted  as  a  non-parametric 
estimate  of  l/gfe).  Since 

it  is  intuitively  obvious  that 

-£  In  nPe(Dn  ();;))  (1) 
ni  =  i 

will  be  an  approximation  of  the  Kullback-Leibler  distance,  I(g,fQ),  defined  by 
I(g,fG)=  Jg(x)(lng(x)-lnf0(x))dx. 

The  approximation  above  is  analogous  to  the  univariate  approximation  given  in 
Ranneby  (1984)  and  the  Maximum  Spacing  Estimate  of  0  is  obtained  by  maximizing 
approximation  (1)  above. 

For  each  class  we  estimate  the  mixture  proportion,  the  mean  vectors  and  the 
covariance  matrices.  Remote  sensing  images  have  the  property  that  the  interclass 
distances  are  rather  small  so  the  different  parts  of  the  mixture  distributions  are  not 
easy  to  separate.  This  makes  it  difficult  to  find  good  starting  values  which  are  crucial 
for  other  estimation  methods.  Although  the  multivariate  MSP-  method  theoretically 
give  an  attractive  estimation  method  much  work  remains  until  all  practical  problems 
are  solved. 

Taxt  et  al  (1991)  found  a  substantial  increase  in  the  correct  classification  rates  when 
using  mixture  models  instead  of  multivariate  normal  distributions. 

CLASSIFICATION  RESULTS 

In  Flygare  (1995)  the  performance  of  some  contextual  classification  methods  is 
evaluated  on  Landsat  TM  data.  Hasletf  s  model  was  used  for  the  conditional  class 
distribution  and  in  three  of  the  methods  she  used  models  with  autocorrelated  spectral 
reflectances.  Methods  utilizing  images  from  one  and  two  occasions  were  compared. 
There  was  no  surprise  that  an  autocorrelation  model  utilizing  information  from  two 
occasions  gave  the  best  classification  results.  Mean  Absolute  Deviation  was  used  to 
compare  the  classification  results  with  the  evaluation  data.  A  21%  improvement  was 
obtained  with  the  two  image  autocorrelation  method  compared  to  the  best  one  image 
method. 


457 


BIOGRAPHICAL  SKETCH 


Bo  Ranneby  is  professor  at  the  Swedish  University  of  Agricultural  Sciences.  He 
took  a  Ph.D.  in  mathematical  statistics  at  Umea  University  in  1975.  In  1992  he  was 
elected  as  an  ordinary  member  of  the  International  Statistical  Institute.  He  is  head  of 
the  Department  of  Forest  Resource  Management  and  Geomatics. 

REFERENCES 

Cheng,  R.C.H.  &  Amin,  N.A.K.  (1983).  Estimating  parameters  in  continuous 

univariate  distributions  with  a  shifted  origin.  J.R.  Statist.  Soc.  B.  45,  394-403. 
Flygare,  A-M.  (1993).  Contextual  classification  using  multi-temporal  Landsat  TM 

data.  Statistical  Research  Report,  University  of  Umea. 
Flygare,  A-M.  (1995).  Comparing  some  contextual  classification  methods  using 

Landsat  TM.  Research  report  no  7,  Institute  of  Mathematical  Statistics,  Umea 

University. 

Flygare,  A-M.  (1996).  An  alternative  estimation  method  for  spatial  autocorrelation 

parameters.  Institute  of  Mathematical  Statistics,  Umea  University. 
Haslett,  J.  (1985).  Maximum  likelihood  discriminant  analysis  on  the  plane  using  a 

Markovian  model  of  spatial  context.  Pattern  Recognition,  18,  287-296. 
Hjort,  N.L.  and  Mohn,  E.  (1984).  A  comparison  of  some  contextual  methods  in 

remote  sensing  classification.  In:  Proceedings  of  the  Eighteenth  International 

Symposium  on  Remote  Sensing  of  the  Environment.  Centre  National  d' Etudes 

Spatiales,  Paris,  1693-1702. 
Matern,  B.  (1979).  The  analysis  of  ecological  maps  as  mosaics.  In:  Spatial  and 

Temporal  Analysis  in  Ecology,  edited  by  R.M.  Cormack  &  J.K.  Ord,  Statistical 

Ecology  Series,  no  8,  271-288. 
Owen,  A.  (1984).  A  neighborhood-based  classifier  for  Landsat  data.  Canadian  Journal 

of  Statistics,  12,  191-200. 
Pielou,  E.C.  (1964).  The  spatial  pattern  of  two-phase  patchworks  of  vegetation, 

Biometrics,  20,  156-167. 
Ranneby,  B.  (1982).  Stochastic  models  of  variation  in  time  and  space.  In:  Statistics  in 

Theory  and  Practice.  Essays  in  Honour  of  Bertil  Matern,  B.  Ranneby,  ed.  Swedish 

University  of  Agricultural  Sciences,  Umea,  Sweden,  227-245. 
Ranneby,  B.  (1984).  The  maximum  spacing  method.  An  estimation  method  related  to 

the  maximum  likelihood  method.  Scand.  J.  Statist.  11,  93-1 12. 
Taxt,  T.,  Hjort,  N.L.  &  Eikvil,  L.  (1991).  Statistical  classification  using  a  linear 

mixture  of  two  multinomial  probability  densities.  Statistical  Research  Report,  no  5, 

Institute  of  Mathematics,  University  of  Oslo. 
Titterington,  D.M.  (1985).  Comments  on  "Estimating  parameters  in  continuous 

univariate  distributions  with  a  shifted  origin".  J.R.  Statist.  Soc.  B  47,  115-1 16. 
Switzer,  P.  (1965).  A  random  set  process  in  the  plane  with  a  Markovian  property, 

Ann.  Math.  Stat.,  36,  1859-1863. 


458 


Generalized  Linear  Mixed  Models  for 
Analyzing  Error  in  a  Satellite-based 
Vegetation  Map  of  Utah 

Gretchen  G.  Moisenl,  D.  Richard  Cutler^, 
and  Thomas  C.  Edwards,  Jr.3 

Abstract. — With  the  increasing  demand  for  broad-scale  vegetation  maps 
for  ecosystem  management  and  conservation  planning  comes  the  need 
for  flexible  tools  to  assess  thematic  accuracy  of  these  maps.  In  this  pa- 
per, we  use  a  generalized  linear  mixed  model  (GLMM)  to  explore  the 
relationship  between  thematic  accuracy  in  the  blackbrush  cover-type  of  a 
satellite-based  vegetation  map  of  Utah  and  various  topographical  and 
heterogeneity  components  of  that  map  Because  of  the  difficulty  in  ac- 
cessing many  rugged  areas  of  this  State,  two  strata  were  defined  based  on 
proximity  to  roads.  Vegetation  type  was  recorded  on  heterogeneous  lin- 
ear clusters  of  sample  points  within  the  "off-road"  strata,  and  on  ran- 
domly distributed  sample  points  within  the  "road"  strata  on  selected 
USGS  quadrangle  maps.  A  binary  response  (correctly  classified  /  incor- 
rectly classified)  was  modeled  as  a  function  of  both  fixed  and  random 
effects  accounting  for  spatially  autocorrelated  observations  and  different 
covariance  structures  for  the  random  effects.  The  modeling  exercise  sug- 
gested a  strong  relationship  between  map  error  in  the  blackbrush  cover- 
type  of  the  Colorado  Plateau  of  Utah,  and  stratum,  slope  and  local  hetero- 
geneity.  

INTRODUCTION 

Thematic  accuracy  of  vegetation  cover  maps  derived  from  satellite  imagery  may 
be  related  to  many  factors,  including  elevation,  aspect,  slope,  local  heterogeneity 
and  distance  to  vegetation  boundaries.  Exploring  the  relationship  between  the 
components  of  the  vegetation  classification  model  and  its  uncertainty  is  a  logical 
step  in  an  analysis  of  map  error  that  is  sensitive  to  both  map  use  and  subsequent 
improvements  of  the  map. 

Although  many  new  techniques  are  being  explored  to  address  map  uncertainty, 
generalized  linear  mixed  models  (GLMM's)  have  yet  to  be  applied.  Through  a 
GLMM,  data  from  any  one  of  a  variety  of  continuous  and  discrete  distributions 
can  be  linked  to  a  linear  structure  that  may  contain  both  fixed  and  random  effects. 
GLMMs  can  also  account  for  correlation  among  observations  as  well  as  among 
random  effects  terms  in  the  linear  structure  (Wolfinger  and  O'Connell  1993). 

^  USDA  Forest  Service,  Intermountain  Research  Station,  Ogden,  UT. 
Department  of  Mathematics  and  Statistics,  Utah  State  University,  Logan,  UT. 

USDI  National  Biological  Service,  Utah  Cooperative  Fish  and  Wildlife  Research  Unit,  Utah  State  University,  Logan, 
UT. 


459 


This  flexibility  may  prove  valuable  in  addressing  map  uncertainty  as  well  as  have 
numerous  other  broad-scale  applications. 

In  this  study,  we  use  a  GLMM  to  explore  the  relationship  between  the  error  in 
the  blackbrush  cover- type  of  a  vegetation  map  of  Utah  and  various  topographical 
and  heterogeneity  components  of  that  map. 

DATA 

A  cover-map  of  Utah,  -219,000  km2  in  size,  was  developed  from  a  state-wide 
Landsat  Thematic  Mapper  (TM)  mosaic  created  from  24  scenes  at  30  m  resolution 
(Homer  et  al.  in  press).  A  total  of  38  cover-types  were  modeled.  Modeling  was 
accomplished  using  a  four  step  modeling  approach.  Steps  included:  (1)  the 
creation  of  a  statewide  seamless  mosaic  of  TM  images;  (2)  the  subsetting  of  the 
mosaic  into  3  ecoregions,  the  Basin  and  Range,  Wasatch-Uinta  and  Colorado 
Plateau  (after  Omernik  1987);  (3)  the  association  of  1,758  state-wide  field  training 
sites  to  spectral  classes;  and  (4)  the  use  of  ecological  parameters  based  on 
elevation,  slope,  aspect  and  location  to  further  refine  spectral  classes  representing 
multiple  cover-types  . 

Following  development  of  this  cover-map,  field  data  were  collected  to  assess 
its  thematic  accuracy  (see  Moisen  et  al.  1994  for  design  considerations,  Edwards 
et  al.  1996  for  analysis).  Of  primary  interest  were  estimates  of  by-class  and 
by-ecoregion  accuracy  of  the  map  at  the  base  model  of  one  ha.  A  total  of  100 
7.5-min  quadrangles  were  randomly  selected  roughly  proportional  to  the  area  of 
the  three  ecoregions.  Two  strata  were  identified  on  each  quadrangle  based  on 
proximity  to  roads.  The  "road"  stratum  consisted  of  all  land  within  1  km  of  a 
secondary  or  better  road.  All  other  lands  fell  within  the  "off-road"  stratum.  On 
each  quadrangle,  ten  points  were  randomly  selected  within  the  road  stratum, 
while  ten  were  collected  in  a  randomly  oriented  heterogeneous  linear  cluster 
within  the  off-road  stratum.  Data  were  then  used  to  assess  map  accuracy  based  on 
procedures  outlined  in  Edwards  et  al.  (1996). 

For  this  study,  a  subset  of  the  state-wide  data  comprised  of  the  blackbrush 
cover-type  within  the  Colorado  Plateau  was  modeled  under  a  GLMM.  Data 
consisted  of  96  sample  points  collected  on  two  strata  (road,  off-road)  on  15 
quadrangles.  Anywhere  from  one  and  ten  blackbrush  points  were  available  for 
each  quad/stratum  combination.  Clustered  blackbrush  data  in  the  off-road  stratum 
were  not  necessarily  adjacent  sample  points  because  blackbrush  polygons  were 
often  intermixed  with  other  cover-types  not  considered  in  this  analysis. 

MODEL 

Using  a  logit  link  function,  the  binary  response  (correctly  classified  /  incor- 
rectly classified)  was  modeled  as  a  function  of  both  fixed  and  random  effects 
while  accounting  for  several  covariance  structures  for  random  effects  and  for  spa- 
tially autocorrelated  errors.  For  this  analysis,  the  observations  on  the  96  sample 
points  were  coded  as  1  when  the  mapped  cover-type  agreed  with  the  ground 


460 


cover-type,  and  as  0  when  they  did  not  agree.  Define  y  to  be  our  data  vector  of  96 
Os  and  Is  satisfying 

y  =  n+e.  (1) 

We  used  a  logit  link  function 

S(|i)-log{|*/(l-|i)}  (2) 

and  modeled 

gOi)-Xp  +  Zv.  (3) 
Here,  p  is  a  vector  of  unknown  fixed  effects  with  known  model  matrix  X,  and 
v  is  a  vector  of  unknown  random  effects  with  known  model  matrix  Z.  Assume 
E(v)  =  0  and  cov(v)  =  G,  where  G  is  unknown.  An  effect  may  be  considered 
fixed  if  the  inference  space  is  limited  to  the  observed  levels  of  that  effect.  An  ef- 
fect may  be  considered  random  if  the  inference  space  is  applied  to  a  population  of 
levels,  not  all  of  which  are  observed.  Fixed  effects  considered  in  this  application 
included  both  discrete  and  continuous  variables.  Because  quadrangle  maps  were 
randomly  selected  for  subsampling  from  a  population  of  quadrangles,  quadrangles 
were  modeled  as  random  effects.  Also,  e  is  a  vector  of  unobserved  errors  with 
£(e|u,)  =  0  and 

cov(8|n)  =  R^RR^.  (4) 

Here      is  a  diagonal  matrix  containing  evaluations  at  \i  of  the  variance  function 

V(n)  =  n(l-H).  (5) 
R  and  G  were  modeled  using  covariance  structures  detailed  below. 


Fixed  Effects 


Nine  fixed  effects  variables  were  considered.  Three  were  topographical  vari- 
ables used  in  the  classification  model  itself.  These  were  extracted  from  a  90  m 
Digital  Elevation  Model  and  include  elevation  in  meters  (ELEV),  slope  in  de- 
grees (SLOPE),  and  aspect.  A  transformation  of  aspect  (TRASP),  used  by  Rob- 
erts and  Cooper  (1989),  takes  the  form 

mn  A  nn    1  -cos( aspect-  30) 

TRASP  =  x—r  -.  (6) 

This  transformation  assigns  the  highest  values  to  land  oriented  in  a  north- 
northeast  direction,  the  coolest  and  wettest  orientation  in  Utah. 

In  addition  to  the  three  topographical  variables,  we  considered  four  different 
measures  of  heterogeneity  surrounding  the  sample  point.  Richness  (RICH)  is  de- 
fined as  the  number  of  cover-types  found  in  the  surrounding  8  pixels.  The  other 
three  heterogeneity  variables,  eveness  (EVEN)  and  two  measures  of  diversity  (Dj 
and  D2),  are  defined  in  table  1.  Higher  values  for  all  indices  indicate  increasing 
heterogeneity. 


461 


Table  1. — Measures  of  heterogeneity.  Here  S  equals  richness,  n  equals  8  pixels,  and  is 
the  number  of  pixels  belonging  to  the  ith  of  S  cover-types.  


Variable 


Source 


Formula 


Hill  (1973) 


H-.2[(?H?)I 


i  =  1 
S  -1 


00 


D2 


EVEN 


Simpson  (1949) 


Ludwig  and  Reynolds  (1988) 


Li  =  1  J 

EVEN  =  (D2-l)/(Drl) 


(8) 
(9) 


Two  other  fixed  effects  considered  were  the  minimum  distance  in  meters  to  a 
different  map  cover-type  (DIST)  and  a  variable  indicating  membership  in  the  road 
or  off-road  stratum  (STRATA).  Strata  was  the  only  categorical  fixed  effect  vari- 
able. All  others  were  continuous. 

Covariance  Structures 

Because  quadrangle  maps  (QUAD)  were  randomly  selected  for  subsampling 
from  a  the  statewide  population  of  quadrangles,  these  were  included  as  random 
effects  in  the  GLMM.  Three  covariance  models  for  G  were  considered  (table  2). 

Table  2. — Covariance  structures  for  G  and  R. 


Structure 


Form 


Simple 

Compound  symmetry 
Varying  coefficients 


=  a  for  i  -  j,  else  0 

2     2  2 
G|j  =  aj  +  a    for  (i  =  j),  else  ax 

2 

Gy  =  o{.  for  (i=j),  else  0 


Spherical  spatial  R..  =  ac2[l-(3dij/2p)  +  (dfj/2p3)]  for  ^  %  p?  ^  Q 


(10) 

(11) 
(12) 

(13) 


A  spherical  spatial  covariance  structure,  illustrated  in  the  last  row  of  table  2,  was 
considered  for  R.  Here  covariance  between  sample  points  is  modeled  as  a  func- 
tion of  distance  between  those  points,  accounting  for  both  correlation  between 
clustered  locations  and  potential  correlation  between  sample  points  in  relatively 
close  quadrangles.  Although  numerous  spatial  structures  could  have  been  tried, 
the  spherical  structure  is  very  flexible  and  converged  more  readily  in  preliminary 
trials. 


462 


Model  Fitting  Strategy 


Parameters  in  the  GLMM  were  estimated  through  pseudo-likelihood  proce- 
dures as  described  in  Wolfinger  and  O'Connell  (1993)  using  a  SAS  macro  sup- 
plied by  Russ  Wolfinger  of  SAS  Institute  Inc.  This  macro  uses  PROC  MIXED 
and  the  Output  Delivery  System,  requiring  SAS/STAT  and  SAS/IML  release  6.08 
or  later. 

An  iterative  model  fitting  strategy  was  adopted.  After  identifying  quadrangle 
maps  as  our  random  effects,  all  fixed  effects  were  included  in  the  model.  We 
tried  all  combinations  of  covariance  structures  for  G  as  listed  in  table  2.  The  best 
covariance  structure  was  selected  based  on  Akaike's  Information  Criterion  and 
Schwarz's  Bayesian  Criterion  (Wolfinger  1993).  In  subsequent  iterations  fixed 
effects  were  dropped  based  on  significance  of  parameter  estimates,  likelihood  ra- 
tio tests,  and  predictive  capability  of  the  model.  Again,  best  covariance  structure 
was  selected  for  each  iteration.  Because  of  collinearity,  the  four  measures  of  het- 
erogeneity were  considered  in  the  model  separately. 

RESULTS 

Likelihood  ratio  tests  and  parameter  significance  levels  led  us  to  favor  a  parsi- 
monious model  containing  only  STRATA,  SLOPE  and  D2  as  fixed  effects.  Ex- 
clusion of  other  variables  had  little  impact  on  the  predictive  capability  of  the 
model  based  on  confusion  matrices  and  plots  of  predicted  values  from  different 
model  trials.  The  covariance  structures  selected  were  simple  and  spherical  for  G 
and  R,  respectively.  The  signs  of  the  fixed  effects  parameters  indicate  the  rela- 
tionship between  error  and  the  variables  (table  3).  In  this  case,  positive  values  for 
SLOPE  and  D2  indicate  that  error  increases  as  SLOPE  and  D2  increase  (figures 
la  -b).  In  contrast,  the  negative  value  for  the  off-road  stratum  indicates  that  prob- 
ability of  error  is  less  in  the  off-road  stratum  and  greater  in  the  road  stratum.  The 
estimate  of  p,  the  spatial  covariance  parameter  in  R,  suggests  that  spatial  depen- 
dence is  negligible  between  sample  units  greater  than  322  meters  apart  (figure 
lc). 


Table  3. — Parameter  estimates  and  their  standard  errors  for  final  GLMM. 


Parameter 

Estimate 

SE 

Pr  >   x    or  (Z  ) 

Intercept 

-  1.78 

0.85 

0.04 

STRATA  (off-road) 

-  1.39 

0.59 

0.02 

SLOPE 

0.12 

0.07 

0.08 

D2 

0.58 

0.31 

0.06 

P 

322.46 

124.86 

0.01* 

a2 

1.63 

1.37 

0.23* 

463 


100         150         200         250         300  350 

Distance  (m) 


Figure  la-b. — Relationship  between  probability  of  map  error  and  fixed  effects, 
lc. — Correlation  modeled  as  a  function  of  distance  between  sample  points. 


DISCUSSION 

Frequently  in  broad-scale  sample  surveys  clustered  and  subsampling  sample 
designs  are  adopted  for  the  sake  of  efficiency  in  estimation  of  population  means 
and  totals.  However,  such  cost-effective  designs  can  hamper  efforts  to  further  ex- 
plore ecological  relationships  under  a  classical  linear  model  framework  by  violat- 
ing model  assumptions  like  independence  and  normality  of  errors.  In  this  study 
we  illustrated  how  a  GLMM  affords  the  flexibility  to  analyze  sample  survey  ac- 
curacy data.  Through  a  GLMM  we  were  able  to  include  both  fixed  and  random 
effects,  account  for  spatially  autocorrelated  errors,  and  allow  for  a  variety  of  cova- 
riance  structures  for  the  random  effects. 

The  model  we  fit  for  blackbrush  contained  some  information  for  map  users  be- 
yond simple  probability  of  misclassification  by  cover-type,  and  also  provided  in- 
formation to  the  map-maker  for  model  improvements.  We  learned  that  incorrect 
mapping  of  this  cover-type  tended  to  occur  near  roads  in  steep  and  heterogeneous 
areas.  The  fact  that  strata  was  a  significant  contributor  to  our  model  of  map  error 
could  be  an  indication  that  vegetation  away  from  roads  differs  from  that  near 
roads,  and  is  governed  by  an  environmental  factors  not  accounted  for  in  the  other 


464 


fixed  effects.  However,  the  better  performance  in  the  off-road  strata  could  be  an 
artifact  of  different  quality  in  data  collection  efforts  between  the  two  strata.  For 
example,  georeferencing  data  in  linear  clusters  may  have  been  easier  or  done  with 
greater  care  ,  making  off-road  data  less  subject  to  positional  error. 

The  inclusion  of  slope  in  our  model  might  highlight  the  difficulty  of  classifica- 
tion in  steep  and  often  shadowy  areas.  Slope  was  not  included  in  the  initial  clas- 
sification model  for  blackbrush  in  the  Colorado  Plateau,  and  its  inclusion  might 
improve  maps  of  that  cover-type.  The  notion  that  classification  in  heterogeneous 
areas  is  tougher  than  in  homogenous  areas  is  not  new,  but  our  model  helps  deter- 
mine the  magnitude  of  heterogeneity's  contribution  to  error.  Also,  numerous  indi- 
ces of  heterogeneity  are  available  in  the  literature  and  we  illustrated  that  some  in- 
dices may  be  make  a  more  significant  contribution  than  others  to  models  of  map 
error.  Map  users  might  also  find  the  model  results  helpful,  making  them  more 
skeptical  of  mapped  blackbrush  on  steeper  slopes  and  in  more  heterogeneous  ar- 
eas. 

ACKNOWLEDGMENTS 

We  would  like  to  thank  our  reviewers  for  their  thoughtful  comments.  We  are 
also  grateful  to  Ron  Tymcio  who  tackled  many  GIS  challenges,  David  Early  for 
data  collection,  Collin  Homer  who  served  as  our  patient  and  ever-present  practical 
conscience,  and  especially  Scott  Bassett,  who  can  program  anything,  fast. 

REFERENCES 

Edwards,  T.  C,  Jr.;  Moisen,  G.  G.;  Cutler,  D.  R.  1996.  Assessing  map  accuracy 
in  an  ecoregion-scale  cover  map.  In  review,  Ecological  Applications. 

Hill,  M.  O.  1973.  Diversity  and  evenness:  a  unifying  notation  and  its  conse- 
quences. Ecology.   54:  427-432. 

Homer,  C;  Ramsey,  R.  D.;  Edwards,  T.  C,  Jr.;  Falconer,  A.  In  press.  Land- 
scape cover-type  modelling  using  a  multi-scene  TM  mosaic.  Photogram- 
metric  Engineering  and  Remote  Sensing. 

Ludwig,  J.  A.;  Reynolds  J.  F.  1988.  Statistical  ecology.  New  York:  John  Wiley 
and  Sons.  337  p. 

Moisen,  G.  G.;  Edwards,  T.  C.  Jr.;  Cutler,  D.  R.  1994.  Spatial  sampling  to  as- 
sess classification  accuracy  of  remotely  sensed  data.  In:  Michener,  W. 
K.;  Brunt,  J.  W.;  Stafford,  S.  G.,  eds.  Environmental  Information  Man- 
agement and  Analysis:  Ecosystem  to  Global  Scales.  London:  Taylor  and 
Francis:  159-176. 

Omernik,  J.  M.  1987.  Map  supplement:  ecoregions  of  the  conterminous  Unitd 
States.  Annals  of  the  Association  of  American  Geographers.  77:  118-125 
(map). 

Roberts,  D.  W.;  Cooper,  S.  V.  1989.  Concepts  and  techniques  of  vegetation 
mapping.  In:  Ferguson,  D.;  Morgan,  P.;  Johnson,  F.  D.,  eds.  Land  Clas- 
sifications based  on  vegetation:  applications  for  resource  management. 


465 


Gen.  Tech.  Rep.  INT-257.  Ogden,  UT:  U.  S.  Department  of  Agriculture, 

Forest  Service,  Intermountain  Research  Station:  90-96. 
Simpson.  E.  H.  1949.  Measurement  of  diversity.  Nature.  163:  688. 
Wolfinger,  R.  D.  1993.  Covariance  structure  selection  in  general  mixed  models. 

Communications  in  Statistics:  Simulation  and  Computation  22(4):  1079- 

1106. 

Wolfinger,  R.  D.;  O'Connell,  M.  1993.  Generalized  linear  mixed  models:  a 
pseudo-likelihood  approach.  Journal  of  Statistical  Computation  and  Simu- 
lation .  48:  233-243. 

BIOGRAPHICAL  SKETCH 

Gretchen  Moisen  is  a  Research  Forester  with  the  Interior  West  Resource  In- 
ventory and  Monitoring  Unit  of  the  Forest  Service's  Intermountain  Research  Sta- 
tion. She  is  currently  working  toward  her  Ph.D.  in  Statistics  at  Utah  State  Univer- 
sity. In  addition  to  her  work  on  map  accuracy,  Gretchen  is  involved  in  sample  de- 
sign and  broad-scale  vegetation  modeling  studies. 

Richard  Cutler  is  an  Associate  Professor  in  the  Department  of  Mathematics 
and  Statistics  at  Utah  State  University.  He  received  his  Ph.D.  in  Statistics  from 
University  of  California,  Berkeley  in  1989  and  his  research  interests  include 
sample  survey  theory,  census  adjustment,  generalized  linear  models  and  robust 
estimation. 

Tom  Edwards  is  a  Research  Ecologist  with  the  National  Biological  Service's 
Utah  Cooperative  Fish  and  Wildlife  Research  Unit,  and  is  an  Assistant  Professor 
in  the  Department  of  Fisheries  and  Wildlife,  Utah  State  University.  Tom  received 
his  Ph.D.  in  wildlife  biology  from  the  University  of  Florida  in  1987.  His  research 
interests  include  wildlife  habitat  relations  modeling  and  the  development  of  meth- 
ods for  assessing  and  monitoring  biological  diversity  at  large  landscape  scales. 


466 


Statistical  Properties  of  Measures  of 
Association  and  the  Kappa  Statistic 

for  Assessing  the  Accuracy  of 
Remotely  Sensed  Data  Using  Double 

Sampling 

Mohammed  A.  Kalkhan1,  Robin  M.  Reich2,  and 
Raymond  L.  Czaplewski3 

Abstract.— A  Monte  Carlo  simulation  was  used  to  evaluate  the  statistical 
properties  of  measures  of  association  and  the  Kappa  statistic  under  double 
sampling  with  replacement.  Three  error  matrices  representing  three  levels 
of  classification  accuracy  of  Landsat  TM  Data  consisting  of  four  forest 
cover  types  in  North  Carolina.  The  overall  accuracy  of  the  five  indices 
ranged  from  0.35%  to  82.1%  depending  on  the  number  of  classes,  the  level 
of  classification  accuracy  of  satellite  imagery,  and  the  simulated  sample 
sizes  of  reference  plots.  Statistical  criteria  used  in  the  evaluation  included: 
percent  bias,  mean  squared  error,  relative  error,  ratio  of  the  mean  variance 
to  the  simulation  variance,  and  95%  confidence  coverage  rates.  Results  of 
the  simulation  indicated  that  double  sampling  provided  unbiased  estimates 
of  the  overall  accuracy  of  remotely  sensed  imagery  irrespective  of  the 
number  of  classes  in  the  image  being  analyzed,  or  sample  size.  While  no 
one  index  was  superior  for  all  levels  of  accuracy,  number  of  classes,  or 
sample  size,  the  Kappa  statistic  and  Pearson's  P  provided  the  best  estimates 
of  the  overall  accuracy  of  the  remotely  sensed  images.  Results  from 
previous  studies  suggest  that  increasing  the  sample  size,  or  reducing  the 
number  of  classification  in  the  remotely  sensed  image  may  increase  the 
accuracy  and  precision  of  the  estimates.  However,  this  was  not  the  case  in 
this  study  because  the  low  accuracy  of  the  aerial  photos  to  ground  data. 


Biometrician.  Natural  Resource  Ecology  Laboratory,  Colorado  State  University.  Fort  Collins,  CO  80523,  USA. 
Associate  Professor,  Department  of  Forest  Sciences,  Colorado  State  University,  Fort  Collins,  CO  80523,  USA. 
USDA  Forest  Sen  ice  Research,  Rocky  Mountain  Forest  and  Experiment  Station.  Fort  Collins.  CO  80526,  USA. 


467 


INTRODUCTION 


High  resolution  satellite  data  (HRSD),  such  as  Landsat  Thematic  Mapper  (TM) 
Data,  the  Systeme  Probatorie  D'Observation  de  la  Terre  (SPOT- 1,2:  France)  satellite 
imagery,  and  JERS-l,OPS  (Japan)  provide  natural  resource  managers  with  vital  information 
for  mapping,  monitoring,  and  assessing  the  condition  of  their  resources  (Kalkhan  1994). 
Manager  are  also  concerned  with  assessing  the  accuracy  of  their  thematic  maps.  The  high 
cost  of  collecting  ground  data  limits  the  sample  sizes  used  in  assessing  the  accuracy  of  a 
thematic  map.  To  reduce  this  cost,  one  could  use  aerial  photographs.  Aerial  photographs  are 
less  expensive  than  ground  data,  but  their  degree  of  misclassification  tends  to  be  higher 
(Kalkhan  et  al.  1995b).  As  an  alternative,  some  combination  of  aerial  photography  and 
ground  data  (i.e.,  double  sampling)  could  be  used  to  improve  the  accuracy  assessment  of  the 
remotely  sensed  data  at  a  lower  cost. 


Measures  of  Association  and  the  Kappa  Statistic 

Indices  developed  by  Turk  (1979),  Hellden  (1980),  and  Short  (1982)  estimate  the 
accuracy  of  individual  categories  of  remotely  sensed  data  based  on  the  errors  of  commission 
and  omission  associated  with  the  estimated  error  matrix.  A  major  disadvantage  of  these 
indices  is  that  they  tend  to  underestimate  the  accuracy  (Rosenfield  and  Fitzpatrick-Lins 
1986).  Another  disadvantage  is  that  no  valid  variance  formula  has  been  developed  for  these 
indices,  making  it  difficult  to  evaluate  the  precision  of  the  estimates  (Kalkhan  1994). 
Because  of  this,  these  indices  are  not  often  used  to  assess  the  accuracy  of  the  remotely 
sensed  images.  In  addition,  users  of  remotely  sensed  images  are  more  interested  in 
evaluating  the  overall  accuracy  of  their  thematic  map  not  just  individual  categories 
(Kalkhan  1994).  As  an  alternative,  users  of  remotely  sensed  images  may  prefer  to  use 
indices  such  as  Pearson's  coefficient  of  mean  square  contingency  (Q2),  Pearson's  P, 
Tschuprow's  T,  or  Cramer's  V,  for  assessing  accuracy.  These  latter  indices  are  related  to  the 
Chi-square  statistic  and  have  more  desirable  statistical  properties  than  the  other  indices 
(Kalkhan  1994).  In  addition,  Bishop  et  al.  (1975,  p  385-386),  Kalkhan  (1994),  and  Kalkhan 
et  al.  (1995a)  provide  detailed  information  about  the  measures  of  association  and 
evaluations  of  their  statistical  properties. 

The  Kappa  statistic  (K)  was  originally  developed  by  Cohen  (1960)  to  measure  the 
observed  agreement  for  categorical  data  (Landis  and  KocK  1977).  Since  then,  the  Kappa 
statistic  has  received  considerable  attention  in  remote  sensing  applications  (Congalton  and 
Mead  1983,  Rosenfield  and  Fitzpatrick-Lins  1986,  Hudson  and  Ram  1987,  Stehman  1992, 
Czaplewski  1994,  Kalkhan  1994).  The  Kappa  statistic  was  described  in  detail  by  Bishop 
et  al.  (1975,  p.  395-400)  and  Kalkhan  (1994). 


468 


Double  Sampling  and  Accuracy  Assessments 

Kalkhan  et  al.  (1995b)  point  out  that  double  sampling  and  accuracy  assessments 
involve  three  phases:  (1)  unbiased  sampling  of  the  landscape;  (2)  accuracy  assessments 
using  a  composite  estimator;  and  (3)  the  integration  of  data  and  accuracy  assessments  into 
a  geographic  information  systems  (GIS).  In  this  paper  we  discuss  phase  2. 

Estimating  the  accuracy  of  maps  derived  from  remotely  sensed  data  requires 
statistical  sampling  of  aerial  photographs  or  ground  reference  plots  to  insure  reliable 
estimates  (Card  1982).  However,  collecting  information  to  assess  the  accuracy  of  a  thematic 
map  with  respect  to  reference  data  is  time  consuming  and  costly.  In  designing  a  cost- 
effective  sampling  method,  two  feats  must  be  accomplished,  both  of  which  affect  the  quality 
of  information  contained  in  the  sample  and  the  precision  of  the  estimate.  First,  an  unbiased 
sample  must  be  selected  from  the  population.  Second,  the  amount  of  variation  in  the  sample 
statistic  must  be  controlled  somewhat  by  the  sampling  method  (Scheaffer  et  al.  1990).  Since 
field  observations  are  expensive,  a  design  that  provides  a  precise  estimate  of  a  parameter 
at  minimum  cost  is  desirable  (Cochran  1977). 

Two  phases  are  required  in  double  sampling.  The  first  phase  is  to  locate  a  sample 
point  on  an  aerial  photograph.  It  is  preferable  to  pinpoint  a  location  in  the  center  of  the 
aerial  photo  because  it  has  less  error  and  distortion  (Kalkhan  et  al.  1995b).  Next,  the 
sampling  point  is  georeferenced  to  the  Landsat  TM  imagery  through  x,y  coordinates.  In  the 
second  phase,  the  point  selected  from  the  aerial  photo  is  located  on  the  ground  to  verify 
cover  type  mapping  accuracy. 

Accuracy  Assessment:  The  Composite  Estimator 

Kalkhan  et  al.  (1995b)  point  out  that  validating  the  accuracy  of  coarse-  and  fine- 
scale  vegetation  maps  using  double  sampling  requires  accurate  estimates  of  bias  and 
variance  at  multiple  spatial  scales.  For  this,  a  composite  estimator  developed  by  Mayback 
(1979)  can  be  used.  The  composite  estimator  is  based  on  data  from  two  phases  of  sampling, 
i.e.  remotely  sensed  data  and  field  data.  Czaplewski  (1992)  proposed  to  use  the  composite 
estimator  as  a  statistical  method  to  improve  the  precision  of  the  estimates.  The  composite 
estimator  provides  considerable  flexibility  to  accommodate  complex  sampling  designs  for 
cross-classified  census  and  sample  survey  data  (Czaplewski  1995).  Adding  aerial 
photography  as  an  intermediate  phase  creates  the  "double  sampling."  More  details  about  the 
composite  estimator  and  its  uses  with  double  sampling  were  provided  by  Kalkhan  et  al. 
(1995b). 

In  assessing  the  classification  accuracy  of  Landsat  TM  Data,  we  used  an  error 
matrix  "contingency  table"  which  is  the  most  common  way  to  represent  the  classification 
accuracy  of  remotely  sensed  data  or  thematic  maps.  The  error  matrix  consists  of  a  series  of 
rows  and  columns  containing  the  number  of  sample  units,  such  as  pixels.  Each  unit  is 
assigned  to  a  particular  category  relative  to  its  actual  type  based  on  a  set  of  reference  data. 
The  error  matrix  provides  the  user  with  information  on  the  accuracy  of  individual 
categories,  and  both  errors  of  commission  and  omission  in  the  classification  (Rosenfield  and 


469 


Fitzpatrick-Lins  1986).  Errors  of  commission  relate  to  user  accuracies,  while  errors  of 
omission  represent  the  accuracy  of  the  remotely  sensed  data. 

The  objective  of  this  study  is  to  evaluate  the  statistical  properties  of  Pearson's 
coefficient  of  mean  square  contingency  (Q2),  Pearson's  P,  Tschuprow's  T,  Cramer's  V,and 
the  Kappa  statistic  for  assessing  the  classification  accuracy  of  remotely  sensed  data  using 
double  sampling. 

METHODS 

The  data  used  in  this  study  were  from  a  pilot  study  designed  to  evaluate  the  use  of 
Landsat  TM  data  in  identifying  forest  cover  types  in  the  state  of  North  Carolina  (Kalkhan 
1994).  A  subsample  of  data  was  used  to  construct  three  composite  error  matrices 
representing  three  levels  of  classification  accuracy  (poor,  average,  and  good)  with  four 
forest  cover  types.  Forest  cover  types  were  identified  using  an  unsupervised  classification 
procedure  available  in  ERDAS  Software  using  bands  TM3  (red),  TM4  (IR1),  and  TM5 
(IR2).  Two  error  matrices  were  required  for  assessing  the  accuracy  of  thematic  maps  using 
the  Kappa  statistic  and  measures  of  association  with  double  sampling.  The  first  error  matrix 
contained  information  from  Landsat  TM  data  and  aerial  photos,  while  the  second  error 
matrix  contained  information  from  aerial  photos  and  ground  data.  Each  error  matrix 
contained  count  data  (i.e.  pixel,  spatial  information)  that  were  converted  to  a  joint 
probability  matrix  for  the  purpose  of  sampling  and  evaluating  the  accuracy  of  the  maps 
(Kalkhan  1994).  Each  error  matrix  was  estimated  using  Monte  Carlo  simulation  (Rubinstein 
1981)  using  double  sampling  with  replacement.  The  error  matrix  relating  the  aerial  photo 
to  ground  data  was  estimated  using  sample  sizes  of  50,  100,  150,  300,  and  800.  The  error 
matrix  relating  the  remotely  sensed  imagery  to  the  aerial  photo  was  estimated  using  sample 
sizes  ranged  from  200  to  25600  depending  on  number  ground  data  used  in  the  second  phase 
of  sampling.  This  process  was  repeated  5000  times  using  a  Monte  Carlo  simulation  based 
on  a  GA  USS  matrix  language  computer  program.  The  bias  for  the  five  indices  was  computed 
as  the  difference  between  the  average  estimates  of  the  measures  of  association  and  the 
Kappa  statistic  and  their  true  population  mean  which  is  known  without  error  in  the  Monte 
Carlo  simulation. 

In  addition  to  knowing  whether  an  estimate  is  unbiased,  it  is  desirable  to  know 
something  about  the  precision  of  the  estimates.  One  method  of  comparing  the  precision  of 
estimates  with  different  amounts  of  bias  is  to  use  the  mean  squared  error  (MSE).  Cochran 
(1977)  defines  mean  square  error  as  the  "Variance  +  Bias2".  Therefore,  an  estimator  with 
a  smaller  MSE  is  considered  more  precise  than  one  with  a  larger  MSE,  even  though  the 
latter  may  have  a  smaller  variance  (Reich  and  Hussin  1993). 

Because  of  the  possible  effect  of  bias  on  the  observed  coverage  rates,  one  can 
estimate  the  impact  of  this  bias  by  examining  the  relative  error.  The  relative  error  is  defined 
as  the  ratio  of  bias  to  the  square  root  of  the  mean  square  error.  A  large  relative  error  has  the 
effect  of  distorting  confidence  probabilities  (Cochran  1977).  For  example,  the  actual 
confidence  level  associated  with  a  nominal  95%  confidence  interval  for  an  estimate  with  a 


470 


relative  error  of  0.20  is  approximately  0.9454  (Reich  and  Hussin  1993).  Absolute  relative 
errors  less  than  0.15  have  little  impact  on  coverage  rates. 

In  addition  to  a  bias  in  estimating  the  overall  accuracy,  there  may  also  be  a  bias 
associated  with  estimating  the  variance.  This  can  affect  the  observed  coverage  rates.  One 
way  of  evaluating  this  bias  is  to  compute  the  ratio  of  the  mean  variance  to  the  simulation 
variance.  The  mean  variance  was  computed  by  averaging  the  5000  estimated  variances 
using  the  classical  variance  formulae  found  in  the  literature  (see  Kalkhan  1994).  For 
estimating  the  variance  of  the  Kappa  statistic,  we  used  Czaplewski's  derived  variance 
approximation  formula  (Czaplewski  1994).  The  simulation  variance  was  computed  as  the 
variance  of  the  5000  simulated  estimates  of  overall  accuracy.  A  95%  confidence  interval  for 
each  index  was  computed  using  the  appropriate  standard  error  of  the  estimate,  and  the 
proportion  of  confidence  intervals  enclosing  the  true  population  mean  was  determined. 

RESULTS  AND  DISCUSSION 

The  observed  (true)  values  of  overall  accuracy  for  the  remotely  sensed  images 
ranged  from  0.33  to  0.87,  while  the  overall  accuracy  of  aerial  photo  to  ground  was  0.79 
(Table  1).  The  observed  overall  accuracy  for  the  Kappa  statistic  and  measures  of  association 
ranged  from  0.004  to  2.03. 

The  scenes  with  four  classes  had  a  mean  accuracy  for  the  Kappa  statistic  ranging 
from  -0.8  to  64.3,  while  Pearson's  Q2,  had  an  overall  accuracy  ranging  from  0.9  to  140.1 
(Table  2).  Pearson's  P  ranged  from  9.4  to  75.7,  while  Tschuprow's  T  and  Cramer's  V  had  an 
overall  accuracy  ranging  from  5.5  to  68. 

Note  that  the  overall  accuracy  of  the  Kappa  statistic  increased  for  poor  classification 
as  the  aerial  photo  to  ground  sample  sizes  increased,  while  for  the  measures  of  association 
their  overall  accuracy  decreased  as  the  sample  sizes  of  remotely  sensed  images  increased. 
For  average  and  good  classifications  the  overall  accuracy  decreased  for  all  indices  when  the 
aerial  photo  to  ground  sample  sizes  increased.  One  could  expect  that  the  overall  accuracy 
of  the  aerial  photo  to  ground  would  be  higher  compared  to  the  accuracy  of  remotely  sensed 
images.  The  Landsat  TM  imagery  was  taken  on  October  8,  1985,  while  the  aerial  photos 
were  taken  on  October  10  and  November  9  of  1985.  The  ground  survey  to  collect  ground 
sampling  points  was  accomplished  during  November  11,  1982  through  November  1986. 
This  might  have  an  effect  on  assessing  the  accuracy.  Increasing  the  sample  size  did  not 
improve  the  overall  accuracy.  In  fact,  the  overall  accuracy  decreased  as  sample  size 
increased. 


Relative  bias 

A  consistent  trend  was  observed  in  the  relative  bias  with  respect  to  the  accuracy  of 
the  scene  and  number  of  samples  at  different  phases  of  sampling  (Table  3).  Pearson's  P 
provided  the  best  estimate  of  overall  accuracy  ranging  from  -0.1  to  -0.8  (Table  3)  followed 
by  Tschuprow's  T,  Cramer's  V,  and  Pearson's  Q2  with  relative  bias  ranging  from  -0.2  to  -1 .0. 


471 


Table  l.--Observed  values  of  five  accuracy  assessment  indices . 


UfluL. 

Poor 

XTirmt 

OCX* 

Ptsto-otoaca 

0 

0.3321 

0.5594 

0.8744 

0.7884 

K 

0.0035 

0.3571 

0 .8206 

0.6967 

Q 

0.2308 

0.9447 

2 .0320 

1.4862 

P 

0.4330 

0.6970 

0. 8186 

0.7732 

T 

0.2774 

0.5611 

0. 8230 

0.7039 

V 

0.3774 

0.5611 

0.8230 

0.7039 

1  Overall  accuracy  (O).  Kappa  jtauoicOC).  Pearson's  cnrtUnm  of 
mean  contingency  (Q1).  Pearson's  (P).  TschuproVs  CD.  and  Cramer's  (V). 


Table  2.— Estimated1  values  of  five  accuracy  assessment  indices. 

Index1  ■  Poor  lTiciffi  0oo4 


K 

50 

-0 

7 

-0 

S 

-0 

7 

-0 

8A- 

30 

3 

30 

3 

30 

0 

29 

6 

64 

1 

64 

0 

64 

3 

64 

2 

Q 

14 

5 

13 

6 

13 

3 

13 

3 

44 

8 

44 

8 

43 

3 

42 

4 

139 

0 

138 

8 

140 

1 

139 

9 

P 

34 

3 

33 

4 

33 

1 

32 

9 

54 

7 

54 

7 

54 

1 

53 

8 

75 

6 

75 

S 

75 

7 

75 

7 

T 

21 

4 

20 

a 

20 

S 

20 

4 

38 

4 

31 

4 

37 

7 

37 

4 

67 

7 

67 

6 

67 

9 

67 

9 

V 

21 

6 

20 

9 

20 

7 

20 

6 

38 

( 

38 

6 

37 

9 

37 

7 

67 

8 

«7 

7 

68 

0 

68 

0 

It 

100 

-0 

4 

-0 

4 

-0 

C 

-0 

5B» 

21 

4 

28 

1 

28 

1 

28 

3 

62 

3 

63 

3 

63 

0 

62 

3 

0 

7 

0 

6 

7 

6 

7 

6 

5 

34 

5 

33 

9 

33 

6 

33 

7 

12S 

4 

12  5 

8 

125 

12  6 

1 

P 

24 

8 

24 

3 

24 

2 

23 

9 

50 

1 

49 

8 

49 

7 

49 

7 

74 

3 

74 

3 

74 

74 

3 

T 

14 

9 

14 

C 

14 

S 

14 

3 

33 

7 

33 

4 

33 

3 

33 

3 

64 

4 

64 

S 

64 

64 

6 

V 

14 

9 

14 

6 

14 

5 

14 

3 

33 

7 

33 

4 

33 

3 

33 

3 

64 

4 

64 

S 

64 

64 

6 

K 

ISO 

-0 

3 

-0 

4 

-0 

3 

-0 

4C 

27 

9 

37 

7 

37 

6 

37 

6 

61 

S 

61 

3 

61 

61 

4 

0 

4 

S 

4 

4 

4 

4 

4 

3 

31 

0 

30 

5 

30 

3 

30 

3 

120 

4 

119 

6 

119 

120 

3 

P 

20 

4 

20 

0 

19 

9 

19 

7 

40 

3 

48 

0 

47 

9 

47 

8 

73 

S 

73 

S 

73 

6 

73 

6 

T 

12 

1 

11 

8 

11 

7 

11 

6 

32 

0 

31 

7 

31 

6 

31 

5 

63 

2 

63 

0 

63 

0 

63 

1 

V 

12 

1 

11 

8 

11 

7 

11 

6 

33 

0 

31 

7 

31 

6 

31 

S 

63 

2 

63 

0 

63 

0 

63 

1 

X 

300 

-0 

3 

-0 

2 

-0 

2 

-0 

2D« 

27 

5 

37 

3 

37 

3 

27 

3 

60 

3 

60 

3 

60 

3 

60 

1 

Q 

3 

3 

2 

2 

2 

2 

2 

1 

37 

9 

37 

5 

27 

4 

37 

3 

113 

1 

112 

9 

113 

0 

112 

3 

P 

14 

S 

14 

3 

14 

1 

14 

1 

46 

5 

46 

3 

46 

3 

46 

3 

72 

7 

72 

7 

72 

7 

72 

6 

T 

8 

t 

i 

4 

a 

2 

• 

2 

30 

4 

30 

2 

30 

2 

30 

1 

61 

3 

61 

2 

61 

3 

61 

1 

V 

8 

S 

e 

4 

* 

2 

8 

2 

30 

4 

30 

3 

30 

3 

30 

1 

61 

3 

61 

2 

61 

3 

61 

1 

X 

eoo 

-0 

1 

-0 

1 

-0 

1 

-0 

0E* 

27 

3 

37 

3 

37 

3 

37 

3 

59 

6 

59 

5 

59 

5 

59 

4 

0 

1 

0 

i 

0 

l 

0 

0 

9 

26 

2 

26 

0 

35 

9 

26 

0 

108 

C 

108 

1 

108 

2 

107 

8 

p 

9 

9 

9 

7 

9 

5 

9 

4 

45 

5 

45 

4 

45 

3 

45 

3 

73 

1 

72 

0 

73 

0 

73 

0 

T 

5 

7 

5 

6 

5 

s 

5 

5 

29 

5 

29 

4 

39 

4 

29 

4 

60 

1 

60 

0 

60 

0 

59 

9 

V 

5 

7 

5 

6 

5 

5 

5 

5 

?9 

5 

29 

4 

29 

4 

29 

4 

60 

1 

60 

0 

60 

0 

59 

9 

1  Estimated  values  were  muJupled  by  100. 
1  Sec  Table  1  for  description  of  variables. 

N  Sample  size  from  aerial  photo  to  ground  (Second  phase  of  samptmf ). 
A  Sample  sizes  from  remotely  sensed  to  serial  photo  are  200. 400.800, 1600. 
B  Sample  sizes  from  remotely  sensed  to  aerial  photo  are  400.100.1600.3200. 
C  Sample  Sizes  from  remotely  sensed  to  serial  photo  are  600. 1 200.2400.4800. 
D  Sample  sizes  from  remotely  sensed  to  aerial  photo  are  121XI.24O0.48OO.9600. 
E  Sample  sizes  from  remotely  sensed  to  aerial  photo  are  3200.6400. 1 2800.25600. 
*  Frist  phase  of  ssmpling  from  remotely  sensed  imagr  to  aerial  photo. 


Table  3.--Percent  bias*  for  five  accuracy  indices. 


Poor 

ATtr»a« 

Oooa 

X 

50 

-3 

1 

-3 

4 

-3 

0 

-3 

3A 

-0 

2 

-0 

2 

-0 

2 

-0 

2 

-0 

2 

-0 

2 

-0 

2 

-0 

2 

Q 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

-0 

5 

-0 

5 

-0 

5 

-0 

6 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

P 

-0 

3 

-0 

2 

-0 

3 

-0 

3 

-0 

3 

-0 

2 

-0 

3 

-0 

3 

-0 

1 

-0 

1 

-0 

1 

-0 

1 

T 

-0 

2 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

2 

-0 

2 

-0 

2 

V 

-0 

2 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

2 

-0 

2 

K 

100 

-2 

1 

-3 

3 

-3 

8 

-2 

4B 

-0 

3 

-0 

2 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

2 

-0 

2 

Q 

-0 

7 

-0 

7 

-0 

7 

-0 

7 

-0 

6 

-0 

6 

-0 

6 

-0 

6 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

P 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

1 

-0 

1 

-0 

1 

-0 

1 

T 

-0 

5 

-0 

5 

-0 

5 

-0 

5 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

-0 

3 

-0 

3 

-0 

2 

-0 

2 

V 

-0 

5 

-0 

5 

-0 

5 

-0 

S 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

-0 

3 

-0 

2 

-0 

2 

-0 

2 

X 

150 

-1 

8 

-3 

0 

-2 

0 

-3 

3C 

-0 

3 

-0 

2 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

Q 

-0 

( 

-0 

8 

-0 

8 

-0 

8 

-0 

7 

-0 

7 

-0 

7 

-0 

7 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

P 

-0 

s 

-0 

5 

-0 

5 

-0 

s 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

1 

-0 

1 

-0 

1 

-0 

1 

T 

-0 

6 

-0 

6 

-0 

6 

-0 

c 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

-0 

2 

-0 

2 

-0 

2 

-0 

2 

V 

-0 

6 

-0 

6 

-0 

6 

-0 

t 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

-0 

3 

-0 

2 

-0 

2 

-0 

2 

K 

300 

-1 

S 

-1 

6 

-1 

6 

-1 

5D 

-0 

2 

-0 

2 

-0 

3 

-0 

2 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

0 

-0 

9 

-0 

9 

-0 

9 

-0 

9 

-0 

7 

-0 

7 

-0 

7 

-0 

7 

-0 

4 

-0 

4 

-0 

4 

-0 

4 

p 

-0 

1 

-0 

7 

-0 

7 

-0 

7 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

1 

-0 

1 

-0 

1 

-0 

1 

T 

-0 

7 

-0 

7 

-0 

7 

-0 

7 

-0 

5 

-0 

5 

-0 

5 

-0 

S 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

V 

-0 

7 

-0 

7 

-0 

7 

-0 

7 

-0 

5 

-0 

s 

-0 

5 

-0 

5 

-0 

3 

-0 

3 

-0 

3 

-0 

i 

X 

800 

-1 

2 

-1 

3 

-1 

1 

-1 

IE 

-0 

2 

-0 

2 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

Q 

-1 

0 

-1 

0 

-1 

0 

-1 

0 

-0 

7 

-0 

7 

-0 

7 

-0 

7 

-0 

5 

-0 

S 

-0 

5 

-0 

5 

P 

-0 

8 

-0 

8 

-0 

8 

-0 

8 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

-0 

1 

-0 

1 

-0 

1 

-0 

1 

T 

-0 

8 

-0 

8 

-0 

8 

-0 

8 

-0 

5 

-0 

5 

-0 

S 

-0 

5 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

V 

-0 

8 

-0 

8 

-0 

8 

-0 

8 

-0 

5 

-0 

5 

-0 

5 

-0 

5 

-0 

3 

-0 

3 

-0 

3 

-0 

3 

Table  5. -Nominal*  95%  confidence  coverage  rates  using  the 
classical  variance  formula  for  five  accuracy  indices. 


laai 


nTUMr 


-2222- 


50 

200 

41 

80 

49 

15 

73 

3 

66 

<• 

1 

10 

15 

12 

19 

4 

12 

400 

26 

66 

29 

* 

46 

1 

50 

2 

0 

4 

9 

6 

11 

3 

18 

800 

17 

50 

18 

c 

29 

1 

36 

1 

0 

2 

7 

4 

7 

3 

11 

1600 

11 

36 

11 

3 

18 

0 

25 

0 

0 

1 

4 

3 

4 

3 

7 

100 

400 

2 

61 

3 

0 

9 

0 

45 

0 

0 

0 

1 

1 

3 

0 

4 

800 

1 

67 

1 

0 

3 

0 

28 

0 

0 

0 

1 

0 

1 

0 

2 

1600 

1 

51 

1 

0 

2 

0 

IB 

0 

0 

0 

0 

0 

0 

0 

0 

3200 

0 

39 

0 

0 

1 

0 

13 

0 

0 

0 

0 

0 

0 

0 

0 

150 

600 

0 

82 

0 

0 

1 

0 

30 

0 

0 

0 

0 

0 

0 

0 

0 

1200 

0 

«7 

0 

0 

0 

0 

16 

0 

0 

0 

0 

0 

0 

0 

0 

2400 

0 

52 

0 

0 

0 

0 

5 

0 

0 

0 

0 

0 

0 

0 

0 

4800 

0 

i'l 

0 

0 

0 

0 

0 

0 

0 

0 

0 

12 

0 

0 

0 

300 

1200 

0 

67 

0 

0 

0 

0 

2 

0 

0 

0 

0 

0 

0 

0 

0 

2400 

0 

54 

0 

0 

0 

0 

1 

0 

0 

0 

0 

0 

0 

0 

0 

4800 

0 

39 

0 

0 

0 

0 

1 

0 

0 

0 

0 

0 

0 

0 

0 

9600 

0 

7 

0 

0 

0 

0 

0 

0' 

0 

0 

0 

82 

0 

0 

0 

800 

3200 

0 

67 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

6400 

0 

53 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

12800 

0 

40 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

25600 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

'  95*  Confidence  Interval  multiply  by  100. 
See  Table  1  for  description  of  variables. 
Sampling  from  remote  sening  image  to  aerial  photo. 


Table  6. --Relative  error*  for  five  accuracy  assessment  indices . 


Poor 

Arareaa 

Oootl 

K 

50  -0 

0  -0 

0 

-0 

0 

-0 

OA 

-3 

6 

-3 

6 

-3 

6 

-3 

6 

-8 

2 

-8.2 

-8 

2 

-8 

2 

Q 

3  -2 

3 

-2 

3 

-2 

3 

-9 

4 

-9 

4 

-9 

4 

-9 

4 

-20 

3 

-20.3 

-20 

3 

-20 

3 

P 

3  -4 

3 

-4 

3 

-4 

3 

-7 

0 

-7 

0 

-7 

0 

-7 

0 

-8 

2 

-a.; 

-8 

2 

-8 

2 

T 

8  -2 

6 

-2 

8 

-2 

8 

-5 

6 

-5 

6 

-5 

6 

-5 

6 

-8 

2 

-8.2 

-8 

2 

-8 

2 

V 

8  -3 

8 

-2 

8 

-3 

8 

-5 

6 

-5 

6 

-5 

6 

-5 

6 

-8 

2 

-8.; 

-8 

2 

-8 

2 

K 

100  -0 

0  -0 

0 

-0 

0 

-0 

0E 

-3 

6 

-3 

6 

-3 

6 

-3 

t 

-8 

2 

-s . : 

-8 

2 

-8 

2 

Q 

3  -3 

3 

-2 

3 

-2 

3 

-9 

4 

-9 

4 

-9 

4 

-9 

4 

-20 

3 

-20.3 

-20 

3 

-20 

3 

P 

3  -4 

3 

-4 

3 

-4 

3 

-7 

0 

-7 

0 

-7 

0 

-7 

0 

-8 

2 

-8.2 

-8 

2 

-8 

3 

T 

8  -2 

8 

-2 

8 

-3 

8 

-5 

6 

-5 

6 

-5 

6 

-5 

6 

-8 

2 

-8.2 

-8 

2 

-8 

2 

V 

B  -2 

8 

-2 

8 

-3 

8 

-5 

6 

-5 

6 

-5 

6 

-5 

6 

-8 

2 

-8.2 

-8 

2 

-8 

3 

K 

150  -0 

0  -0 

0 

-0 

0 

-0 

OC 

-3 

6 

-3 

6 

-3 

6 

-3 

6 

-8 

2 

-8  . ; 

-8 

2 

-8 

2 

c 

3  -2 

3 

-2 

3 

-3 

3 

-9 

4 

-9 

4 

-9 

4 

-9 

4 

-20 

3 

-20.: 

-20 

3 

-20 

3 

p 

3  -4 

3 

-4 

3 

-4 

3 

-7 

0 

-7 

0 

-7 

0 

-7 

0 

-8 

2 

-8.; 

-8 

2 

-8 

2 

T 

8  -3 

8 

-2 

8 

-2 

8 

-5 

6 

-5 

6 

-5 

6 

-5 

6 

-8 

2 

-8.2 

-8 

2 

-8 

2 

V 

8  -3 

8 

-2 

8 

-2 

8 

-5 

6 

-5 

6 

-5 

6 

-5 

S 

-8 

2 

-8  .. 

-8 

2 

-8 

2 

K 

300  -0 

0  -0 

0 

-0 

0 

-0 

0D 

-3 

6 

-3 

6 

-3 

6 

-3 

6 

-8 

2 

-8  . : 

-8 

2 

-8 

3 

Q 

3  -3 

3 

-2 

3 

-2 

3 

-9 

4 

-9 

4 

-9 

4 

-9 

4 

-20 

3 

-20.. 

-20 

3 

-20 

3 

P 

3  -4 

3 

-4 

3 

-4 

3 

-7 

0 

-7 

0 

-7 

0 

-7 

0 

-8 

2 

-8  . . 

-8 

2 

-8 

3 

T 

8  -2 

8 

-2 

8 

-2 

8 

-S 

6 

-5 

6 

-S 

t 

-5 

6 

-8 

2 

-8.. 

-8 

2 

-8 

2 

V 

B  -2 

8 

-2 

8 

-2 

8 

-5 

6 

-5 

6 

-S 

6 

-5 

6 

-8 

2 

-8.. 

-8 

.2 

-8 

2 

K 

800  -0 

0  -0 

0 

-0 

0 

-0 

0E 

-3 

6 

-3 

f 

-3 

6 

-3 

6 

-8 

2 

-8.. 

-8 

.2 

-8 

2 

Q 

3  -2 

3 

-2 

3 

-2 

3 

-9 

4 

-9 

4 

-9 

4 

-9 

4 

-20 

3 

-20. 

-20 

.3 

-20 

.3 

P 

.3  -4 

3 

-4 

3 

-4 

.3 

-7 

0 

-7 

0 

-7 

0 

-7 

0 

-8 

.2 

-8.. 

-8 

.2 

-8 

.2 

T 

8  -2 

8 

-2 

8 

-2 

8 

-5 

< 

-5 

6 

-5 

6 

-5 

.6 

-8 

2 

-8.. 

1  -8 

.2 

-8 

2 

V 

8  -2 

8 

-2 

8 

-2 

8 

-5 

6 

-5 

6 

-5 

6 

-5 

6 

-8 

2 

-8. 

1  -8 

.2 

-8 

2 

*  Percent  Mas  divided  by  100. 

1  See  Tsbie  1  and  2  for  description  of  variables  and  sample  mzes. 


•  Relative  error  rrtiltipry  by  1000. 

'  See  Table  1  and  2  for  description  of  variables  and  sample  sizes. 


Table  7. -Ratio  of  mean  variance  to  simulation  variance  for  five 
accuracy  indices. 


Table  4.--Mean  square  error*  of  five  accuracy  indices. 

Index1 


X 

50 

954 

3 

1179 

8 

894 

8 

1077 

7A 

2 

3 

2 

3 

2 

6 

1 

0 

■1 

8 

f 

8 

1 

1 

4 

7 

0 

13 

8 

16 

7 

17 

9 

18 

3 

37 

7 

37 

7 

29 

5 

30 

3 

10 

0 

10 

.1 

9 

6 

9 

7 

p 

4 

3 

5 

3 

5 

6 

5 

8 

4 

6 

4 

6 

s 

0 

S 

3 

0 

6 

0 

.6 

0 

S 

0 

6 

T 

S 

2 

6 

3 

6 

7 

6 

9 

10 

0 

10 

0 

10 

8 

11 

1 

3 

3 

3 

3 

3 

1 

3 

1 

V 

5 

0 

6 

1 

6 

5 

6 

7 

9 

7 

9 

7 

10 

s 

10 

8 

3 

1 

3 

1 

3 

0 

3 

0 

X 

100 

441 

6 

500 

0 

789 

1 

561 

3B 

4 

1 

4 

5 

4 

s 

4 

S 

5 

9 

S 

a 

6 

0 

S 

8 

Q 

48 

3 

50 

1 

50 

3 

51 

7 

40 

3 

41 

1 

41 

s 

41 

4 

14 

7 

14 

5 

14 

7 

14 

4 

p 

18 

2 

19 

3 

19 

4 

20 

1 

7 

9 

8 

1 

8 

3 

8 

2 

0 

9 

0 

.9 

0 

9 

0 

8 

T 

21 

4 

33 

6 

22 

7 

33 

6 

16 

0 

16 

4 

16 

6 

16 

S 

4 

7 

4 

7 

4 

7 

4 

6 

V 

21 

4 

22 

6 

22 

7 

33 

6 

16 

0 

16 

4 

16 

6 

16 

5 

4 

7 

4 

7 

4 

7 

4 

6 

K 

150 

319 

3 

412 

3 

392 

3 

468 

9C 

4 

B 

S 

1 

S 

2 

5 

3 

C 

3 

6 

4 

C 

4 

6 

4 

Q 

64 

0 

65 

3 

65 

5 

66 

3 

45 

1 

45 

9 

46 

1 

46 

4 

16 

6 

16 

.9 

16 

9 

16 

7 

P 

28 

0 

29 

1 

29 

3 

29 

8 

9 

4 

9 

7 

9 

0 

9 

9 

1 

0 

1 

0 

1 

0 

1 

0 

T 

31 

9 

33 

0 

33 

3 

33 

■ 

18 

s 

18 

9 

19 

0 

19 

3 

S 

4 

s 

5 

S 

5 

5 

4 

V 

31 

9 

33 

0 

33 

3 

33 

8 

18 

5 

18 

9 

19 

0 

19 

3 

5 

4 

5 

5 

s 

5 

5 

4 

K 

300 

225 

3 

273 

0 

257 

1 

238 

8D 

5 

3 

5 

5 

5 

5 

5 

5 

7 

0 

7 

1 

7 

1 

7 

2 

Q 

80 

9 

81 

s 

82 

2 

83 

1 

49 

6 

50 

3 

SO 

4 

SO 

5 

19 

6 

19 

8 

19 

7 

20 

0 

P 

43 

* 

44 

7 

45 

4 

45 

6 

11 

0 

11 

3 

11 

3 

11 

4 

1 

3 

1 

3 

1 

3 

1 

3 

T 

47 

8 

48 

7 

49 

4 

49 

5 

21 

0 

21 

3 

31 

4 

31 

5 

6 

S 

6 

5 

6 

5 

6 

< 

V 

47 

8 

48 

7 

49 

4 

49 

5 

31 

0 

31 

3 

21 

4 

31 

5 

6 

s 

< 

5 

6 

5 

C 

6 

X 

800 

144 

1 

167 

7 

131 

4 

130 

3E 

5 

5 

S 

C 

5 

7 

5 

C 

7 

s 

7 

5 

7 

S 

7 

c 

Q 

91 

3 

91 

5 

91 

8 

93 

0 

S3 

3 

52 

5 

52 

7 

53 

6 

21 

7 

21 

9 

21 

9 

22 

1 

P 

59 

t 

60 

2 

60 

9 

61 

1 

12 

1 

12 

3 

13 

3 

12 

3 

1 

4 

1 

4 

1 

4 

1 

5 

T 

63 

0 

63 

S 

64 

1 

64 

4 

22 

S 

22 

6 

33 

7 

23 

7 

7 

3 

7 

3 

7 

3 

7 

4 

V 

63 

0 

63 

5 

64 

1 

64 

4 

22 

5 

22 

6 

22 

7 

22 

7 

7 

3 

7 

3 

7 

3 

7 

4 

X 

50 

0 

44 

0 

24 

0 

12 

0 

06A 

0 

34 

0 

34 

0 

10 

0 

05 

0 

24 

0 

12 

0 

06 

0 

03 

Q 

0 

38 

0 

30 

0 

10 

0 

04 

0 

28 

0 

28 

0 

08 

0 

04 

0 

30 

0 

10 

0 

05 

0 

02 

P 

0 

46 

0 

23 

0 

12 

0 

06 

0 

31 

0 

31 

0 

08 

0 

04 

0 

27 

0 

10 

0 

OS 

0 

02 

T 

0 

OS 

0 

03 

0 

01 

0 

01 

0 

03 

0 

03 

0 

01 

0 

00 

0 

03 

0 

01 

0 

01 

0 

00 

V 

0 

97 

0 

49 

0 

25 

0 

12 

0 

62 

0 

62 

0 

17 

0 

09 

0 

68 

0 

23 

0 

11 

0 

06 

X 

100 

0 

48 

0 

23 

0 

11 

0 

06B 

0 

34 

0 

17 

0 

09 

0 

04 

0 

24 

0 

13 

0 

06 

0 

03 

a 

0 

47 

0 

26 

0 

13 

0 

07 

0 

38 

0 

19 

0 

11 

0 

05 

0 

26 

0 

13 

0 

07 

0 

03 

p 

0 

53 

0 

25 

0 

12 

0 

06 

0 

34 

0 

18 

0 

10 

0 

OS 

0 

25 

0 

13 

0 

06 

0 

03 

T 

0 

06 

0 

03 

0 

01 

0 

01 

0 

04 

0 

02 

0 

01 

0 

00 

0 

03 

0 

01 

0 

01 

0 

00 

V 

1 

16 

0 

SS 

0 

26 

0 

14 

0 

76 

0 

38 

0 

21 

0 

10 

0 

55 

0 

28 

0 

14 

0 

07 

X 

150 

0 

48 

0 

26 

0 

13 

0 

07C 

0 

39 

0 

21 

0 

11 

0 

06 

0 

28 

0 

14 

0 

07 

0 

04 

Q 

0 

53 

0 

25 

0 

12 

0 

07 

0 

37 

0 

20 

0 

10 

0 

05 

0 

27 

0 

.14 

0 

07 

0 

03 

P 

0 

57 

0 

28 

0 

14 

0 

08 

0 

37 

0 

20 

0 

10 

0 

05 

0 

27 

0 

14 

0 

07 

0 

03 

T 

0 

06 

0 

03 

0 

01 

0 

01 

0 

04 

0 

02 

0 

01 

0 

01 

0 

03 

0 

.03 

0 

01 

0 

00 

V 

1 

27 

0 

.62 

0 

30 

0 

17 

0 

83 

0 

44 

0 

22 

0 

11 

0 

60 

0 

3  1 

0 

16 

0 

08 

X 

300 

0 

49 

0 

26 

0 

14 

0 

07D 

0 

43 

0 

22 

0 

11 

0 

06 

0 

30 

0 

.15 

0 

08 

0 

04 

Q 

0 

60 

0 

.30 

0 

16 

0 

08 

0 

39 

0 

21 

0 

11 

0 

06 

0 

30 

0 

15 

0 

08 

0 

04 

P 

0 

62 

0 

31 

0 

17 

0 

08 

0 

40 

0 

21 

0 

11 

0 

06 

0 

30 

0 

15 

0 

08 

0 

04 

T 

0 

07 

0 

03 

0 

02 

0 

01 

0 

04 

0 

02 

0 

01 

0 

01 

0 

03 

0 

02 

0 

01 

0 

00 

V 

1 

39 

0 

70 

0 

37 

0 

18 

0 

89 

0 

47 

0 

24 

0 

13 

0 

68 

0 

.34 

0 

18 

0 

09 

X 

800 

0 

SO 

0 

.27 

0 

15 

0 

07E 

0 

43 

0 

23 

0 

12 

0 

06 

0 

31 

0 

16 

0 

08 

0 

04 

Q 

0 

62 

0 

.31 

0 

16 

0 

08 

0 

40 

0 

22 

0 

11 

0 

06 

0 

33 

0 

17 

0 

09 

0 

04 

P 

0 

63 

0 

.32 

0 

16 

0 

09 

0 

40 

0 

22 

0 

11 

0 

06 

0 

33 

0 

.17 

0 

09 

0 

04 

T 

0 

07 

0 

04 

0 

02 

0 

01 

0 

04 

0 

02 

0 

01 

0 

01 

0 

04 

0 

02 

0 

01 

0 

00 

V 

1 

41 

0 

71 

0 

36 

0 

19 

0 

90 

0 

50 

0 

25 

0 

13 

0 

75 

0 

38 

0 

19 

0 

10 

1  See  Tabic  1  and  2  far  description  of  variable!  and  sample  iizei. 


•  MSE  divided  by  100. 

1  See  Table  1  and  2  for  description  of  variables  and  sample  sues. 


472 


The  Kappa  statistic  had  high  relative  bias  ranging  from  -0.2  to  -3.4.  As  the  overall  accuracy 
of  the  scene  increased,  the  relative  bias  decreased  for  all  indices.  The  relative  bias  for  the 
Kappa  statistic  decreased  when  the  aerial  photo  to  ground  sample  sizes  increased  for  poor 
classification.  The  opposite  was  found  for  the  measures  of  association  for  average  and  good 
classification;  the  bias  increased  as  the  sample  size  increased. 

Mean  squared  error 

Pearson's  P  had  the  smallest  MSE  across  all  sample  sizes  ranging  from  0.6  to  61.1 
followed  by  Cramer's  V  and  Tschuprow's  T  with  the  MSE  ranging  from  3.0  to  64.4  for  all 
levels  of  classification  (Table  4).  Pearson's  Q2  had  a  MSE  ranging  from  9.7  to  92,  while  the 
Kappa  statistic  had  the  largest  MSE.  In  general,  the  mean  square  error  decreased  with 
increased  accuracy  of  the  remotely  sensed  image.  The  MSE  for  the  Kappa  statistic 
decreased  as  the  aerial  photo  to  ground  sample  sizes  increased,  while  the  opposite  was 
observed  for  the  measures  of  association.  The  MSE  increased  consistently  for  all  indices  as 
the  sample  size  associated  with  estimating  the  error  matrix  relating  that  to  aerial  photo  to 
the  ground.  For  the  poor  classification,  the  Kappa  statistic  decreased  as  the  sample  size  of 
the  photo-ground  increased. 

95  Percent  coverage  rates 

None  of  the  five  indices  provided  consistent  coverage  rates  across  all  sample  sizes 
and  levels  of  accuracy  (Table  5).  The  Kappa  statistic  provided  reasonable  coverage  rates, 
but  was  not  consistent  for  the  poor  classification.  The  coverage  rate  was  very  poor  for  the 
average  and  good  classification.  The  inconsistency  in  coverage  rates  may  be  due  to  the  bias, 
which  could  distort  the  confidence  intervals.  The  lower  than  expected  coverage  rates  may 
also  be  due  to  a  bias  in  estimating  the  variance. 

Bias  relative  to  the  root  mean  squared  error 

In  computing  the  relative  errors,  the  simulation  variance  was  used  instead  of  the 
classical  variance  to  facilitate  comparison  between  all  accuracy  indices.  Except  for  a  few 
instances,  all  the  absolute  relative  errors  exceeded  0.20  (Table  6).  This  suggests  that  the 
observed  biases  are  large  enough  relative  to  the  variability  to  distort  the  expected  coverage 
rates.  This  would  help  explain  why  the  coverage  rates  were  lower  than  expected. 

Ratio  of  the  mean  variance  to  the  simulation  variance 

As  noted,  the  lower  than  expected  coverage  rates  also  may  be  due  to  a  bias 
associated  with  estimating  the  variance.  One  way  to  evaluate  this  bias  is  to  calculate  the 
ratio  of  the  mean  variance  to  the  simulation  variance  (Table  7).  It  is  assumed  that  the 
simulation  variance  provides  an  unbiased  estimate  of  the  variance.  Thus,  a  ratio  of  less  than 
1  indicates  an  underestimation  of  the  variance,  while  a  ratio  greater  than  1  indicates  an 


473 


overestimation  (Cochran  1977).  Using  this  as  a  guideline,  one  can  see  that  all  indices 
underestimated  the  variance,  except  for  a  few  cases  where  Cramer's  V  overestimated  the 
variance.  There  was  also  some  indication  that  the  bias  decreased  as  the  aerial  photo  to 
ground  sample  sizes  increased  for  poor  classification.  In  addition,  as  the  level  of  the 
accuracy  improved,  the  bias  increased. 

The  biases  associated  with  estimating  the  variances  may  be  due  to  the  asymptotic 
nature  of  the  variance  formulae  which  assume  large  sample  sizes.  When  used  with  a  small 
sample  size  the  variance  formulae  may  not  provide  valid  estimates  of  the  variance  (see 
Kalkhan  1994),  especially  for  the  Kappa  statistic  where  we  used  derived  variance  formula 
developed  by  Czaplewski  (1994).  Also,  the  variance  estimates  of  Pearson's  P,  Tschuprow's 
T,  and  Cramer's  V  are  based  on  a  nonlinear  transformation  of  the  asymptotic  variance  of 
Pearson's  Q2.  This  transformation  may  also  introduce  a  bias  in  the  estimation  process.  In 
general,  increasing  the  sample  size  of  the  remotely  sensed  image  did  not  seem  to  decrease 
the  bias  of  the  mean  variance  estimates.  This  is  because  the  overall  accuracy  of  aerial  photo 
to  ground  comparison  was  less  than  the  Landsat  TM  imagery. 

CONCLUSION 

The  Kappa  statistic  and  Pearson's  P  provided  the  best  estimates  of  overall  accuracy 
for  all  levels  of  classification,  sample  sizes,  and  numbers  of  forest  cover  types.  In  addition, 
the  results  of  the  Kappa  statistic  in  this  study  are  applicable  to  Pielou's  index  of  segregation 
(Pielou  1960,  1977;  Kalkhan  etal.  1995a).  Poor  performance  of  the  measures  of  association 
were  due  to  large  biases  associated  with  the  statistic.  The  ratios  of  the  estimated  mean 
variance  to  the  simulation  variance  suggested  that  the  classical  variance  formula  for  the 
various  indices  provided  biased  estimates  of  variance,  irrespective  of  the  number  of  classes, 
or  the  sample  size.  This  bias  may  be  due  to  the  asymptotic  nature  of  the  variance  formulae. 
These  variance  formulae  may  not  provide  valid  estimates  when  used  with  small  sample 
sizes.  Some  of  the  bias  associated  with  the  variance  may  be  due  to  the  nonlinear 
transformation  used  to  approximate  the  variance  of  Pearson's  P,  Tschuprow's  T,  and 
Cramer's  V.  To  overcome  this  problem,  one  could  use  bootstrapping  (Efron  1979)  to  provide 
unbiased  estimates  of  the  variance  (Kalkhan  1994). 

Double  sampling,  as  applied  in  this  study,  allows  one  to  make  inference  about  the 
whole  image  and  not  a  preselected  subset  of  homogenous  test  sites.  It  also  suggests  that, 
increasing  the  sample  size,  or  reducing  the  number  of  classes  on  the  remotely  sensed  image 
may  increase  the  accuracy  and  precision  of  the  estimates.  However,  this  was  not  the  case 
in  this  study  because  the  accuracy  of  this  particular  aerial  photo  to  ground  was  low.  No 
matter  what  indices  one  uses  to  assess  the  accuracy  of  a  remotely  sensed  image, 
bootstrapping  should  be  used  to  ensure  unbiased  estimates  of  the  variance  (Kalkhan  1994). 
To  improve  the  findings  of  this  study,  we  suggest  the  use  of  a  large  number  of  classes  for 
assessing  the  accuracy  of  the  thematic  map.  Prior  to  assessing  the  accuracy  of  any  thematic 
map,  the  user  of  remotely  sensed  must  be  familiar  with  the  statistical  properties  of  any 
accuracy  assessment  index  before  using  it. 


474 


REFERENCES 


Bishop,  Y.  M.  M.,  S.  E.  Feinberg,  and  P.  W.  Hooland.  1975.  Discrete  multivariate 
analysis-theory  and  practice.  MIT  Press,  Cambridge,  MA.  575  p. 

Card,  D.  H.  1982.  Using  known  map  categorical  marginal  frequencies  to  improve  map 
accuracy.  P hot ogramme trie  Eng.  &  Remote  Sensing.  48:431-439. 

Cochran,  W.  G.  1977.  Sampling  Techniques.  3rd  ed.  John  Wiley  and  Sons,  New  York.  428 

P- 

Cohen,  J.  1960.  A  coefficient  of  agreement  of  nominal  scales.  Educ.  Psychological  Meas. 
20:37-46. 

Congalton,  R.  G.,  and  R.  A.  Mead.  1983.  A  quantitative  method  to  test  for  consistency  and 
correctness  in  photointerpretation.  Photogramm.  Eng.  &  Remote  Sens.  49:67-74. 

Czaplewski,  R.  L.  1992.  Accuracy  assessment  of  remotely  sensed  classification  with  multi- 
phase sampling  and  the  multivarite  composite  estimator.  P.  22  In:  1992  (XVIth) 
Intern.  Biometric  Conference,  Hamilton,  New  Zealand,  December  7-11.  Vol.  2,  268 

P- 

Czaplewski,  R.  L.  1994.  Variance  approximation  for  assessment  of  classification  accuracy. 
Research  Paper,  Rocky  Mountain  For.  &  Range  Exper.  Station,  Forest  Serv. 
U.S.D.A.,  94  p. 

Czaplewski,  R.  L.  1995.  Assessment  of  classification  accuracy  and  extent  estimates  for  land 

cover  map  with  double  sampling.  Forest  Science  (In  Review). 
Efron,  B.  1 979.  Bootstrap  methods:  Another  look  at  the  jackknife.  Annals  of  Stat.  7: 1  -26. 
Helldon,  U.  1980.  A  test  of  Landsat-2  imagery  and  digital  data  for  thematic  mapping 

illustrated  by  anenvironmental  study  in  northern  Kenya.  Swedan,  Lund  University 

Natural  geography  Institute,  Report  No.  7. 
Hudson,  W.  D.,  and  C.  W.  Ram.  1987.  Correct  formulation  of  the  Kappa  coefficient  of 

agreement.  Photogramm.  Eng.  &  Remote  Sens.  53:421-422. 
Landis,  J.  R.,  and  G.  G.  Kock.  1977.  The  measurement  of  observer  agreement  for 

categorical  data:  Biometrics.  33:154-174. 
Kalkhan,  M.  A.  1994.  Statistical  properties  of  six  accuracy  indices  using  simple  random  and 

stratified  random  sampling:  An  application  in  remote  sensing.  Ph.D.  Dissertation. 

Colorado  State  University.  134  p. 
Kalkhan,  M.  A.,  R.  M.  Reich,  and  R.  L.  Czaplweski.  1995a.  Statistical  properties  of  five 

accuracy  indices  in  assessing  the  accuracy  of  remotely  sensed  data  using  simple 

random  sampling.  In:  1995  ACSM/ASPRS  Annual  Convention  &  Exposition, 

ASPRS  Technical  Papers,  February  27-  March,  1995,  Charlotte,  North  Carolina.  P. 

246-257. 

Kalkhan,  M.  A.,T.  J.  Stohlgren,  and  M.  Coughneour.  1995b.  An  investigation  of 
biodiversity  and  Landscape-scale  gape  patterns  using  double  sampling:  A  GIS 
approach.  In:  1995  Ninth  International  Symposium  on  Geographic  Information 
Systems  for  natural  resources,  environment  and  land  information  management, 
March  27-30,  1995,  Vancouver,  British  Columbia,  Canada.  P.  708-712. 


475 


Maybeck,  P.S.  1979.  Stochastic  Models,  Estimation,  and  Control,  vol.  1.  Academic  Press, 
New  York. 

Pielou,  E.  C.  1961.  Segregation  and  symmetry  in  two  species  populations  as  studied  by 

nearest  neighbor  relations.  J.  Ecology.  49:255-269. 
Pielou,  E.  C.  1977.  Mathematical  Ecology.  John  Wiley  and  Sons,  New  York.  385  p. 
Reich,  R.  M.,  and  Y.  A.  Hussin.  1993.  Estimating  average  stand  biomass  for  a  regional 

forest  inventory  using  radar  backscatter.  ITC  Journal.  1:82-87. 
Rosenfield,  G.  H.  and  K.  Fitzpatrick-Lins.  1986.  A  coefficient  of  agreement  as  a  measure 

of    thematic    classification    accuracy.    Photogramm.     Eng.     &  Remote 

Sensing.  52:223-227. 

Rubinstein,  R.  Y.  1981.  Simulation  and  the  monte  carlo  method.  John  Wily  and  Sons,  New 
York.  278  p. 

Scheaffer,  R.  L.,  W.  Mendenhall,  and  L.  Ott.  1990.  Elementary  Survey  Sampling,  4th  ed., 

PWS-Kent  Publishing  Company,  Boston.  MA.  390  p. 
Short,  N.  M.  1982.  The  Landsat  tutorial  workbook:  Basis  of  satellite  remote  sensing. 

Greenbelt,  MD.,  GSF  Center,  NASA  Reference  Publication  1078. 
Stehman,  S.  V.  1992.  Comparison  of  systematic  and  random  sampling  for  estimating  the 

accuracy  of  maps  generated  from  remotely  sensed  data.  Photogramm.  Eng.  & 

Remote  Sens.  58:1343-1350. 
Turk,  G.  1979.  GT  index:  A  measure  of  the  success  of  prediction.  Remote  Sens.  Environ. 

8:65-75. 

BIOGRAPHICAL  SKETCH 

Mohammed  A.  Kalkhan  is  a  Research  Scientist/GIS  Specialist  with  the  Natural 
Resource  Ecology  Laboratory,  Colorado  State  University.  He  graduated  from  University  of 
Mosul,  Iraq  with  B.Sc.  (Forestry)  in  1973  and  M.Sc.  (Forest  Mensuration)  in  1980. 
Mohammed  holds  a  Ph.D.  in  Forest  Biometrics  from  Colorado  State  University  (1994). 
Research  interest  include  using  remote  sensing,  GIS,  biometrics,  and  spatial  statistics  to 
study  natural  resources,  landscape-gap  analysis  and  ecology,  biodiversity,  and  accuracy 
assessment  of  remotely  sensed  data  with  sampling  techniques. 

Robin  M.  Reich  is  an  Associate  Professor  with  the  Department  of  Forest  Sciences, 
Colorado  State  University.  Research  interest  spatial  statistics,  natural  resource  sampling  and 
forest  biometrics.  Current  research  is  aimed  at  integrating  of  spatial  and  geostatistical 
technique  with  GIS  and  remote  sensing  to  model  the  spatial  interaction  biological 
populations  have  with  themselves  and  their  environment. 

Raymond  L.  Czaplewski  has  spent  the  past  13  years  working  on  statistical  design 
and  analysis  of  regional  monitoring  systems  for  forests  and  other  environmental  features, 
primarily  the  Forest  Service's  national  Forest  Inventory  and  Analysis  Program.  Other 
accomplishments  include  statistical  design  of  the  1990  and  2000  global  surveys  of  tropical 
forests  by  the  Food  and  Agricultural  Organization  of  the  United  Nations,  and  a  design  to 
monitor  the  effectiveness  of  the  President's  1994  Forest  Plan  for  the  States  of  Oregon, 
Washington,  and  California.  He  has  specialized  in  linking  sample  surveys  to  remotely 
sensed  data  and  geographic  information  systems. 

476 


Comparing  Methods  used  to 
Combine  Multi-level  Samples 
of  Compositional  Data 

Murray  Todd  Williams1 


Abstract. — Accuracy  assessment  of  categorical  data  with  complex 
sampling  designs  is  an  important  tool  in  the  use  of  thematic  maps  and 
other  applications.  One  important  development  of  note  is  the  composite 
estimator— an  application  of  the  Kalman  filter.  This  paper  reviews  the 
composite  estimator  and  compares  it  with  standard  weighted  least 
squares  regression.  In  many  cases  the  two  methods  produce  identical 
results,  although  there  are  situations  in  which  one  has  advantages  over 
the  other.  The  conclusion  also  raises  some  critical  questions  which  arise 
from  these  classical  methods,  and  it  examines  contemporary  analyses  of 
compositional  data.  In  conjunction  with  the  methodology  described 
here,  a  public-domain  software  application  has  been  developed. 


INTRODUCTION 

Much  of  the  work  discussed  here  emerged  from  natural  resource  issues, 
specifically  the  analysis  of  forest  inventories  carried  out  by  the  USDA  Forest 
Service.  However,  the  application  of  this  work  has  already  proven  to  be  quite 
large.  Let  us  broadly  define  the  sort  of  data  which  can  benefit  from  these 
analyses. 

In  single-level  categorical  data,  measured  quantities  (frequently  discrete)  are 
identified  as  one  of  a  group  of  mutually  exclusive  categories.  One  example  of  this 
is  categorizing  the  type  of  forest  in  a  section  of  land  by  tree  species.  Another 
example  would  be  the  a  typical  blood  test  in  which  the  number  of  specific  compo- 
nents (red  blood  cells,  lymphocytes,  etc.)  are  counted. 

Multi-stage  and  multi-phase  categorical  sampling  uses  multiple  classifiers  to 
categorize  each  sampling  unit.  For  example,  in  the  above  forest  inventory  situa- 
tion, the  composition  of  a  section  of  land  can  be  measured  using  three  different 
methods:  satellite  imaging,  aerial  photography  (photo  interpretation),  and  manual 
observations  in  the  field. 

In  this  example,  let  a  primary  sampling  unit  be  a  section  of  land  with  an  area 
of  47  acres.  Table  1  shows  an  example  where  the  three  identification  methods 
was  used  to  classify  each  acre  of  land.  Notice  that  38  acres  of  land  were  classified 

^Student.  Department  of  Statistics.  Colorado  State  University,  Ft.  Collins,  CO 


477 


the  same  by  each  method.  For  the  other  9  acres,  there  is  some  degree  of  disagree- 
ment between  categories.  (These  situations  are  shaded  in  the  table.)  In  most 
situations  a  high  level  of  agreement  is  considered  optimal. 

Notice  that  Table  1  also  displays  the  quantities  of  each  multiple-classification 
as  proportions  of  the  total.  This  vector  of  proportions  is  referred  to  as  a 
composition. 


Table  1. — Mutli-level  data  in  a  typical  primary  sampling  unit. 


Satellite  Imaging 

Aerial  Photography 

Ground  Surveys 

Acres 

Composition 

Non-forest 

Non-forest 

Non-forest 

12 

0.255 

Non-forest 

Light  forest 

Light  Forest 

2 

0.043 

Non-forest  * 

Light  forest 

Heavy  forest 

5 

0.106 

Light  forest 

Light  forest 

Light  forest 

15 

0.319 

Light  forest 

Heavy  forest 

Light  forest 

2 

0.043 

Heavy  forest 

Heavy  forest 

Heavy  forest 

11 

0.234 

Total: 

47 

1.000 

There  are  many  situations  in  which  this  type  of  data  is  encountered.  Some  of 
the  possible  applications  become  apparent: 

•  Multiple  classification  techniques.  In  these  situations,  various  techniques 
might  have  different  levels  of  accuracy.  Such  accuracy  is  frequently  inversely 
proportional  to  cost.  In  the  above  example,  satellite  data  might  be  readily 
available  and  therefore  inexpensive,  but  the  degree  of  accuracy  obtained  may 
be  fairly  poor.  Ground  surveys  are  considered  to  be  the  most  accurate,  but 
they  are  also  exceedingly  expensive. 

•  Multiple  time  periods.  A  large  database  of  monitoring  information  may  be 
available  for  previous  time  periods  (say  for  example,  a  large  survey  taken  in 
1970  and  1980)  where  data  for  a  recent  time  period  (1990)  is  more  limited.  A 
combination  of  the  larger  two-period  database  may  be  combined  with  the 
small  sample  of  plots  which  have  all  three  time  periods  in  order  to  extrapolate 
a  better  estimate  of  the  forest  composition  in  1990. 


Construction  of  the  Sample  Composition  Estimates 

Consider  an  (n  x p)  data  matrix  X  with  n  independent  primary  sampling  units 
(PSUs)  and  p  distinct  categories.  Note  that  the  category  labels  resemble  those  in 
Table  1 .  Also  let  each  row  of  X  represent  a  composition,  thus  summing  to  one. 
The  calculations  of  the  composition  estimate  Y  (the  sample  mean  of  the  data 
matrix  X)  and  its  sample  co variance  are  straightforward: 

Y  =  ttXt1    and  SY  =  ^XT(I  -  11T)X  [1] 


476 


COMPOSITE  ESTIMATOR 


The  composite  estimator  was  first  applied  to  categorical  analyses  by 
Czaplewski  (1992)  as  an  application  of  the  Kalman  filter.  Instead  of  motivating 
its  derivation  I  will  just  describe  its  application. 

Consider  two  composition  estimates.  The  estimates  may  have  any  dimension- 
ality, but  for  this  example  we  will  define  a  two-dimensional  composition  Y,  and  a 
three-dimensional  composition  Y2.  Y2  would  be  a  vector  with  the  unit-sum 
constraint.  The  elements  of  Y2  will  correspond  with  the  rows  in  Table  1 . 

Yj  will  only  have  categories  defined  on  only  two  classifications:  Satellite 
imaging  and  Aerial  photography.  Refer  to  Table  2.  It  is  necessary  now  to 
construct  a  binary  transformation  matrix  which  defines  the  unique  mapping  from 
Y,  to  Y2.  Take  this  partition  of  the  binary  matrix  and  call  it  H.  An  example  of  H 
is  shown  as  a  subset  of  the  binary  matrix  in  Table  3. 


Table  2. — A  typical  primary  sampling  unit  for  a  two-dimensional  composition  Y,. 


Satellite  Imaging 

Aerial  Photography 

Acres 

Composition 

Non-forest 

Non-forest 

19 

0.268 

Non-forest 

Light  Forest 

8 

0.113 

Light  forest 

Light  forest 

22 

0.310 

Light  forest 

Heavy  forest 

7 

0.099 

Heavy  forest 

Heavy  forest 

15 

0.21  1 

Total: 

71 

1.000 

From  Czaplewski  (1994)  the  composite  estimator  Fcan  be  expressed  as 

Y  =  Y2  +  K(Y,  -  HY2)    and    SY  =  (I  -  KH)E2  [2] 
where     K  =  E2HT(E,  +  H^H1")  [3] 

The  resulting  composite  estimator  will  have  the  same  dimension  and  category 
labels  as  the  largest  component  Y2.  In  a  sense,  the  composite  estimator  can  be 
considered  a  method  of  improving  the  estimate  of  Y2  by  using  another  independ- 
ent and  unbiased  vector  estimate  (F7)  of  lower  dimensionality. 

A  problem  which  can  sometimes  occur  with  both  the  composite  estimator  and 
the  weighted  least  squares  is  that  the  resulting  composition  estimate  Y  can  have 
negative  components.  For  the  composite  estimator,  there  is  a  partial  remedy  of 
this  problem.  Multiply  K  in  the  above  equations  with  a  scalar  constant  a. 

K  =  a22HT(S1  +  HE2HT)  [4] 

In  optimal  situations  (where  negative  components  do  not  result)  we  can  let 
<x=l.  If  negative  components  of  the  composite  estimator  are  observed,  we  can 
replace  a  with  the  largest  possible  value  in  the  interval  [0,1]  such  that  the  result- 
ing composite  estimator  is  well-behaved. 


479 


WEIGHTED  LEAST  SQUARES 


The  use  of  weighted  (generalized)  least  squares  allows  a  greater  level  of  flexi- 
bility than  the  composite  estimator.  The  only  thing  about  the  weighted  least 
squares  regression  that  may  not  be  considered  straightforward  is  the  fact  that  the 
covariance  matrices  of  the  compositions  will  be  necessarily  singular.  By  replac- 
ing the  standard  matrix  inverse  with  the  generalized  inverse,  the  problem  is 
alleviated. 


Table  3. — Creation  of  the  design  matrix  X. 


•a 

WLS  Regression  Categories  § 

\  i 

\ 

Original  Categories             \  v 

Satellite            Aerial              Ground  a 

Non-forest     Non-forest  Non-forest 
Non-forest     Light  forest    Light  forest 
Non-forest     Light  forest    Heavy  forest 
Light  forest    Light  forest    Light  forest 
Light  forest    Heavy  forest  Light  forest 
Heavy  forest  Heavy  forest  Heavy  forest 

Non-forest 

Non-forest 

H  matrix  for 
the  composites 
estimator. 

t 

1 

0 

0 

0 

0 

0 

Non-forest 

Light  forest 

0 

1 

1 

0 

0 

0 

Light  forest 

Light  forest 

0 

0 

0 

1 

0 

0 

Light  forest 

Heavy  forest 

0 

0 

0 

0 

1 

0 

Heavy  forest 

Heavy  forest 

0 

0 

0 

0 

0 

1 

Non-forest 

Non-forest 

Non-forest 

1 

0 

0 

0 

0 

0 

<missing> 

Light  forest 

Light  forest 

0 

1 

0 

1 

0 

0 

Non-forest 

Light  forest 

Light  forest 

0 

1 

0 

0 

0 

0 

Non-forest 

Light  forest 

Heavy  forest 

0 

0 

1 

0 

0 

0 

Light  forest 

Light  forest 

Light  forest 

0 

0 

0 

1 

0 

0 

Light  forest 

Heavy  forest 

Light  forest 

0 

0 

0 

0 

1 

0 

Heavy  forest 

<missing> 

Heavy  forest 

0 

0 

0 

0 

0 

1 

Heavy  forest 

Heavy  forest 

Heavy  forest 

0 

0 

0 

0 

0 

1 

Let  Yj  and  Y2  be  two  compositions  (unit-sum  vectors)  with  corresponding 
covariance  matrices  27;  and  U2  respectively.  The  two  compositions  can  have  a 
different  dimensionality.  Let  Yj  have  two  classifications  (Satellite  and  Aerial)  and 
let  Y2  have  three  dimensions.  We  can  also  introduce  missing  data  with  weighted 
least  squares.  Hence,  we  will  let  Y2  have  two  missing  categories.  The  categories 
of  these  compositions  are  shown  on  the  left  side  of  Table  3 . 


480 


Let  the  design  matrix  X  be  specified  by  the  binary  matrix  depicted  in  Table  3. 
This  matrix  is  easily  formed  by  placing  a  1  in  each  cell  where  all  the  available 
categories  from  the  original  data  (rows)  match  up  with  their  corresponding  regres- 
sion categories  (columns). 

If  we  consider  F7  and  Y2  as  partitions  of  a  single  vector  Y,  and  define 

y-  y\        and     2  -    o   £2  [5] 

The  generalized  regression  equation  for  the  new  WLS  estimate  (W)  and  its 
covariance  matrix  can  be  given  by 

W  =  (XT2"X)"XTS"Y 

and  [6] 

SW  =  (XTS  ~X)~ 

where  A  is  the  generalized  inverse2  of  A. 

DISCUSSION 

We  have  shown  that  the  composite  estimator  and  weighted  least  squares  are 
both  capable  of  combining  multi-level  composition  estimates  of  different  dimen- 
sionality. Both  methods  require  a  binary  transformation  matrix  which  maps  the 
categories  of  the  corresponding  compositions  to  those  of  the  resulting  estimate. 

Weighted  least  squares  has  an  advantage  over  the  composite  estimator  when 
multiple  estimates  need  to  be  combined.  It  is  also  unique  in  its  ability  to  handle 
data  with  missing  categories.  The  composite  estimator,  however,  is  able  to  work 
with  a  composition  estimate  of  zero  covariance.  (An  example  of  this  is  an  entire 
census  of  a  population,  where  the  population  statistic  represents  no  sampling 
error.)  In  addition,  the  composite  estimator  is  the  only  method  which  can  be 
modified  to  assure  that  no  components  of  the  resulting  estimate  are  negative. 

Inherent  Difficulties  with  Compositional  Data 

The  unit-sum  constraint  is  an  intrinsic  property  of  any  composition.  Little,  if 
any,  attention  has  been  paid  to  this  restriction  despite  the  fact  that  it  greatly 
complicates  any  analysis  of  correlation  between  categories.  Karl  Pearson  (1897) 


By  definition.  A   is  the  generalized  inverse  of  A  if:  A=AA  A  and  A   -A  A  A  .  See  Graybill  (1976) 


481 


first  pointed  out  difficulties  inherent  with  the  interpretation  of  correlations 
between  ratios  whose  numerators  and  denominators  contain  common  parts. 

Since  that  time,  difficulties  which  have  been  encountered  while  trying  to  inter- 
pret these  correlations  have  been  described  in  papers  by  Chayes  (1960,1962), 
Krumbein  (1962),  Moisimann  (1962),  and  in  books  by  Chayes  (1971)  and  Le 
Maitre(1982). 

Aitchison  (1986)  points  out  four  basic  difficulties  which  arise  from  the  use  of 
a  standard  co variance  matrix  in  analysis. 

Negative  bias  difficulty 

For  a  D-part  composition,  we  have  the  restriction 

cov(xi^c2)  +  cov(x!^c3)  +  •■••  +  cov(x!,xD)  =  -var(xj)  [7] 

Although  the  correlations  between  all  parts  of  a  composition  would  be 
expected  to  freely  span  the  range  [0,1]  this  restriction  requires  that  at  least  one  of 
the  correlations  assumes  a  negative  value.  Such  forced  negative  correlation  could 
cause  some  misinterpretation  of  the  data. 

Subcomposition  difficulty 

In  most  multivariate  situations,  an  m-part  subset  of  an  ft-part  composition 
would  be  expected  to  preserve  the  covariance  structure.  For  example,  consider  a 
4-part  composition  of  categories  A,  B,  C  and  D,  and  then  a  smaller  subset  of  just 
the  categories  A,  B,  and  C. 

Table  4. — Correlation  matrices  from  a  4-category  sample  and  its  3-category  subsample. 

Correlation  for  {A,B,C,D}  Correlation  for  {A,B,C} 

A  B  C  D  ABC 


A  1.000  0.267  -0.410  -0.484  A 
B  0.267  1.000  "  -0.266  "  -0.419  B 
C      -0.410      -0.266       1.000      -0.490  C 


1.000 

-0.066 

-0.701 

-0.066 

1.000 

-0.665 

-0.701 

-0.665 

1.000 

D      -0.484      -0.419      -0.490  1.000 

The  above  data  comes  from  a  hypothetical  data  set.  The  correlation  matrix  on 
the  left  side  is  taken  by  calculating  all  four  categories.  If  we  instead  consider  the 
composition  of  only  the  first  three  categories,  the  resulting  correlation  matrix  is 
given  on  the  right  side.  Notice  the  correlation  between  categories  A  and  B  is 
positive  (0.267),  but  when  removing  the  fourth  category,  the  same  correlation 
between  A  and  B  becomes  negative  (-0.066). 

Basis  difficulty 

The  above  correlation  matrices  were  computed  from  a  multivariate  sample  of 
unit-sum  vectors.  When  we  consider  the  original  data  which  might  be  measured 
(before  the  unit-sum  normalization  is  performed),  we  might  expect  the  sample 


482 


correlation  matrix  of  the  original  data  to  closely  resemble  the  samnle  correlation 
matrix  of  the  normalized  data.  This  is  not  the  case,  and  the  two  correlation  matri- 
ces may  be  dramatically  different.  This  suggests  obvious  problems  when  we  try 
to  use  the  covariance  or  correlation  matrices  of  compositional  data  to  interpret  the 
relationship  between  two  variables. 

Null  correlation  difficulty 

It  is  traditional  to  consider  null-correlation  a  good  indicator  of  independence. 
However,  in  the  situations  we've  already  pointed  out,  a  truly  independent  sample 
will  necessarily  have  negative  correlations  between  elements.  It  is  possible  to 
calculate  the  naturally  occurring  negative  correlations  which  would  occur  under 
an  independence  assumption  and  use  this  to  compare  with  experimental  results, 
but  this  problem  is  also  fraught  with  difficulty. 

The  Contemporary  Analogs  to  a  Composite  Estimator  or  WLS 

In  his  text,  Aitchison  (1986)  provides  an  approach  for  analyzing  the  correla- 
tion structure  of  compositional  data  using  a  log-ratio  approach.  The  natural 
progression  for  this  field  of  complex  sampling  designs  is  to  attempt  to  replace  the 
analysis  of  covariance  matrices  with  these  more  contemporary  methods.  Whether 
it  is  possible  (and  valid)  to  continue  to  use  the  weighted  least  squares  and 
composite  estimator,  applying  the  techniques  of  Aitchison  to  approximations 
derived  from  the  resulting  covariance  matrices,  or  if  these  methods  might  need  to 
be  completely  replaced  by  new  methods  is  currently  unknown. 

ACAS  Software  Application 

The  procedures  described  in  this  paper  have  been  developed  into  a  computer 
package  titled  "ACAS."  (Williams  1995)  This  application  has  been  written  in 
C++  and  is  available  to  the  public.  Although  ACAS  is  a  useable,  full-featured 
package,  continual  development  is  underway  to  provide  more  advanced  tools. 
The  ACAS  system  already  includes  some  of  the  methodology  of  Aitchison  and 
aspires  to  further  develop  contemporary  methods  based  on  his  work. 

ACKNOWLEDGEMENTS 

The  development  of  this  material  would  not  be  possible  without  the  ACAS 
software  project,  which  has  been  mostly  written  by  David  J.  C.  Beach.  This  work 
is  currently  funded  by  the  State  of  Minnesota  Department  of  Natural  Resources 
and  the  USDA  Forest  Service. 


483 


REFERENCES 


Aitchison,  J.  1986.  The  statistical  analysis  of  compositional  data.  Chapman  and 
Hall.  416  p. 

Chayes,  F.  1960.  On  correlation  between  variables  of  constant  sum.  Journal 
Geophysical  Research,  65:4185-4193. 

Chayes,  F.  1962.  Numerical  correlation  and  petrographic  variation.  Journal  of 
Mathematical  Geology,  70:440-452. 

Chayes,  F.  1971.  Ratio  Correlation.  University  of  Chicago  Press. 

Czaplewski,  R.L.  1992.  Accuracy  assessment  of  remotely  sensed  classifications 
with  multi-phase  sampling  and  the  multivariate  composite  estimator.  Hamil- 
ton, New  Zealand.  16th  International  Biometric  Conference.  2:22. 

Czaplewski,  R.L.  1996.  Assessment  of  classification  accuracy  and  extent 
estimates  for  a  land  cover  map  with  double  sample.  Submitted  to  Forest 
Science. 

Graybill,  F.A.  1976.  Theory  and  application  of  the  linear  model.  Wadsworth  & 

Brooks/Cole  Adv.  Books  and  Software.  704p. 
Grizzle,  J.E.,  Starmer,  C.F.,  and  Koch,  G.G.  1969.  Analysis  of  categorical  data  by 

linear  models.  Biometrics  25:489-504. 
Krumbein,  W.C.  1962.  Open  and  closed  number  systems  stratigraphic  mapping. 

Bull.  Amer.  Assoc.  Petrol.  Geologists,  46:  2229-2245. 
Le  Maitre,  R.W.  1982.  Numerical  Petrography.  Amsterdam:  Elsevier. 
Lee,  E.S.,  Forthofer,  R.N.,  and  Lorimor,  R.J.  1989.  Analyzing  complex  survey 

data.  Sage  University  Paper  series  on  Quantitative  Applications  in  the  Social 

Sciences,  07-071.  Beverly  Hills:  Sage  Pubns.  78p. 
Mosimann,  J.E.  1962.  On  the  compound  multinomial  distribution,  the  multivari- 
ate B-distribution  and  correlations  among  proportions.  Biometrika,  49:65-82. 
Pearson,  K.  1897.  Mathematical  contributions  to  the  theory  of  evolution.  On  a 

form  of  spurious  correlation  which  may  arise  when  indices  are  used  in  the 

measurement  of  organs.  Proc.  R.  Soc,  60:489-498. 
Williams,  M.T.,  and  Beach,  D.J.C.    1995.    ACAS  0.4:  accuracy  assessment 

system  program  manual.  USDA  Forest  Service,  Rocky  Mountain  Forest  and 

Range  Experiment  Station,  Fort  Collins,  CO.  33  p.  +  source  code. 

BIOGRAPHICAL  SKETCH 

Murray  Todd  Williams  is  currently  finishing  his  M.S.  degree  in  the  depart- 
ment of  Statistics  at  Colorado  State  University.  The  work  which  appears  in  this 
paper,  along  with  the  development  of  the  ACAS  software  application,  is  a  part  of 
his  thesis.  Murray  received  his  B.A.  in  Mathematics  from  Pomona  College  in 
Claremont,  CA. 


484 


Cost-Effective,  Practical  Sampling 
Strategies  for  Accuracy  Assessment 
of  Large-Area  Thematic  Maps 

Stephen  V.  Stehman* 

Abstract.  —  Accuracy  assessment  is  an  expensive  but 
necessary  process  in  the  development  and  eventual  use  of  a 
large-scale  thematic  map.  The  sampling  strategy  used  for 
collecting  and  analyzing  the  reference  samples  required  for 
the  accuracy  assessment  must  be  cost-effective,  yet  still 
achieve  satisfactory  precision  for  the  estimated  accuracy 
parameters.  Strata  and  clusters  may  be  used  to  improve  the 
efficiency  of  the  sampling  strategy,  and  more  specialized 
designs  such  as  double  sampling  or  adaptive  cluster  sampling 
may  provide  better  precision  for  certain  objectives.  Post- 
stratified  and  regression  estimators  may  be  combined  with  a 
simple  design  to  yield  enhanced  precision  without  sub- 
stantially increasing  costs.  Each  strategy  has  strengths  and 
weaknesses  and  no  single  strategy  is  ideal  for  all  applications. 


INTRODUCTION 

Large-area,  satellite-based  land  cover  mapping  projects  are  becoming  increasingly 
important  for  use  in  environmental  monitoring  and  modeling,  resource  inventories,  and 
global  change  research.  Assessing  the  thematic  accuracy  of  these  maps  requires 
balancing  the  needs  of  statistical  validity  and  rigor  with  the  practical  constraints  and 
limited  resources  available  for  the  task.  Finding  this  balance  is  a  characteristic  of 
practically  any  applied  sampling  problem.  The  primary  expense  is  obtaining  ground 
reference  samples,  the  critical  data  for  an  accuracy  assessment.  A  cost-effective, 
practical  accuracy  assessment  strategy  must  efficiently  use  these  ground  reference 
samples.  This  efficiency  can  be  achieved  by  selecting  an  efficient  sampling  design,  or  by 
using  efficient  analysis  procedures  following  collection  of  the  data. 

A  site-specific,  thematic  accuracy  assessment  is  assumed  of  primary  interest,  with 
the  evaluation  unit  being  a  single  pixel  as  suggested  by  Janssen  and  van  der  Wei  (1994). 
Analogous  strategies  could  be  developed  for  a  polygon-based  accuracy  assessment,  but 
that  topic  is  not  pursued  here.  Each  pixel  on  the  target  map  corresponds  to  a  particular 
location  on  the  ground,  and  it  is  assumed  that  a  ground  visit  leads  to  a  correct  land- 
cover  classification  of  that  pixel.  Location  error  will  be  confounded  with  classification 
error  if  field  site  locations  do  not  correspond  exactly  with  map  locations.  This  view  of 
accuracy  assessment  permits  illustration  of  the  basic  principles  and  techniques  to  be 
presented  in  this  paper. 

The  statistical  approach  taken  is  the  classical  finite  population  sampling  model 
(Cochran  1977).  In  this  perspective  as  applied  to  accuracy  assessment  (Stehman  1995), 
the  population  consists  of  the  N  pixels  on  the  map.  If  a  complete  census  of  ground 
reference  sites  were  obtained,  each  pixel  on  the  map  could  be  labeled  as  correctly  or 

^SUNY  College  of  Environmental  Science  and  Forestry,  320  Bray  Hall,  Syracuse,  NY 


485 


incorrectly  classified,  and  a  population  error  matrix  constructed  from  these  results  in 
which  the  rows  of  the  error  matrix  represent  the  map  classifications,  and  the  columns 
represent  the  reference  classifications  ("truth").  The  objective  of  the  accuracy 
assessment  is  to  use  a  sample  of  reference  locations  to  construct  a  sample  error  matrix 
and  then  to  compute  estimates  of  various  accuracy  parameters,  typically  the  overall 
proportion  of  pixels  correctly  classified  (Pc),  user's  accuracy,  producer's  accuracy,  and 
the  kappa  coefficient  of  agreement  («). 

A  basic  recommendation  is  that  a  probability  sampling  design  should  be  used  to 
collect  the  reference  data.  For  a  probability  sample,  each  pixel  in  the  population  has  a 
positive,  known  probability  of  being  included  in  the  sample.  This  "inclusion 
probability"  is  a  characteristic  of  the  sampling  design.  By  knowing  the  inclusion 
probability  of  each  sampled  pixel,  the  pixels  can  be  correctly  weighted  when  estimating 
accuracy  parameters.  All  pixels  need  not  have  the  same  inclusion  probability,  a 
stratified  sample  being  a  good  example,  as  long  as  the  inclusion  probability  is  known. 
Probability  sampling  provides  an  objective,  scientifically  defensible  protocol  for  accuracy 
assessment. 

In  the  finite  sampling  perspective,  a  sampling  strategy  consists  of  two  parts,  the 
sampling  design,  which  is  a  protocol  by  which  the  sample  data  are  collected,  and  an 
estimator,  which  is  a  formula  for  estimating  a  population  parameter.  An  "efficient" 
sampling  strategy  is  one  in  which  the  variance  of  an  estimator  is  low,  or,  equivalently, 
precision  is  high.  Efficiency  translates  into  cost  savings  in  that  if  a  sampling  strategy 
can  provide  the  same  variance  using  fewer  sampling  locations  than  another  strategy,  the 
former  strategy  is  more  cost-effective.  Efficiency  can  be  gained  by  judicious  choice  of 
design  or  estimator,  or  both.  Stratification,  cluster  sampling,  double  sampling,  and 
adaptive  cluster  sampling  are  design  options  for  improving  efficiency.  Regression  and 
poststratified  estimators  are  estimation  techniques  commonly  employed  in  finite 
population  sampling  to  improve  precision.  These  estimation  techniques  require 
auxiliary  information  in  addition  to  the  sample  reference  data.  Aerial  photography  or 
videography  are  potential  sources  of  auxiliary  information,  and  the  number  of  pixels 
classified  into  each  cover  type  by  the  land-cover  map  can  be  used  as  auxiliary 
information  in  poststratification. 

ENHANCING  PRECISION  VIA  SAMPLING  DESIGN 

Stratification 

Several  stratification  options  exist.  Stratifying  by  geographic  region  ensures  that 
the  sampling  design  for  accuracy  assessment  results  in  a  regionally  well-distributed 
sample,  spreads  the  workload  out  geographically,  and  allows  for  convenient  reporting  of 
accuracy  statistics  by  region.  For  parameters  summarizing  the  full  error  matrix,  such  as 
Pc  and  AC,  geographic  stratification  normally  does  not  produce  large  gains  in  precision 
over  simple  random  sampling  (SRS)  (Cochran  1977,  p.  102)  and  should  not  be  selected 
with  the  hope  of  achieving  a  substantial  reduction  in  variance  for  these  estimators. 

Although  geographic  stratification  is  convenient  for  reporting  accuracy  statistics  by 
region,  it  is  not  necessary  to  stratify  to  obtain  regional  estimates.  Any  geographic 
region  can  be  identified  as  a  subpopulation  and  accuracy  statistics  calculated  for  that 
subpopulation.  The  advantage  of  stratification  is  that  the  sample  size  can  be  controlled 
in  each  region  or  stratum,  and  the  stratum-specific  estimates  will  have  smaller  variance 
compared  to  the  estimates  obtained  via  SRS  or  another  equal  probability  design  such  as 
systematic  or  cluster  sampling.    For  any  equal  probability  design,  the  sample  size  in  a 


486 


region  will  be  proportional  to  the  area  of  that  region,  and  for  small  regions,  this  sample 
size  may  be  too  small  to  achieve  adequate  precision. 

Stratifying  by  the  land-cover  classes  identified  on  the  target  map  has  the  advantage 
of  guaranteeing  a  specified  sample  size  in  each  land-cover  class,  and  thus  is  cost-effective 
if  the  primary  objective  is  estimating  user's  accuracy.  Efficiency  for  estimating 
producer's  accuracy,  Pc,  and  k  may  be  reduced  when  the  design  is  stratified  by  the 
classes  identified  on  the  map  depending  on  the  allocation  of  samples  to  the  strata. 
Stratifying  by  the  map  classification  also  requires  that  the  map  be  completed  prior  to 
selecting  the  sample,  and  this  may  result  in  a  delay  between  the  time  of  imagery  and 
the  time  the  ground  sampling  takes  place.  Changes  in  land  cover  may  occur  during  the 
intervening  period. 

Still  another  possible  stratification  option  was  employed  by  Edwards  et  al.  (1996)  in 
a  large-area  accuracy  assessment  in  Utah,  USA.  Within  broad  geographical  strata,  two 
additional  spatial  strata  were  created,  a  stratum  identified  by  a  1-km  wide  corridor 
centered  on  a  road,  and  an  off-road  stratum.  The  advantage  of  this  stratification 
scheme  is  that  access  to  reference  sites  is  much  easier  for  areas  in  close  proximity  to  a 
road,  and  the  number  of  samples  per  unit  cost  can  be  increased  dramatically  by 
increasing  the  sampling  intensity  in  the  road  stratum.  Selecting  some  samples  from  the 
non-road  stratum  maintains  the  probability  sampling  characteristic  of  the  design. 

Cluster  Sampling 

Cluster  sampling  is  another  potentially  cost-effective  design  for  selecting  reference 
samples.  In  this  design,  the  primary  sampling  unit  (psu)  consists  of  a  cluster  of  pixels, 
such  as  a  3x3  block  or  a  linear  row  of  pixels.  In  one-stage  cluster  sampling,  all  pixels 
within  the  psu  are  sampled,  whereas  in  two-stage  cluster  sampling,  a  subsample  of  pixels 
within  each  psu  is  selected.  The  advantage  of  cluster  sampling  is  that  the  number  of 
pixels  sampled  per  unit  cost  is  increased  because  pixels  are  sampled  in  closer  proximity. 
Moisen  et  al.  (1994)  demonstrated  the  efficiency  gains  achievable  by  cluster  sampling 
based  on  an  analysis  taking  into  account  sampling  costs  and  the  spatial  autocorrelation 
of  classification  errors.  The  pixels  within  a  given  sampled  psu  cannot  be  regarded  as 
independent  observations,  so  the  standard  error  formulas  must  reflect  the  cluster 
structure  of  the  sampling  design  (Czaplewski  1994,  Moisen  et  al.  1994,  Stehman  1996b). 
The  usual  SRS  variance  estimators  do  not  take  into  account  the  within  cluster 
correlation  and  will  likely  underestimate  the  cluster  sampling  variance. 

Sampling  Rare  Land-Cover  Classes 

The  emphasis  placed  on  sampling  rare  land-cover  categories  strongly  influences  the 
design  selected  for  accuracy  assessment.  If  rare  classes  are  considered  extremely 
important,  the  sampling  design  should  reflect  this  priority  and  practically  forces  using  a 
stratified  design  to  ensure  adequate  sample  sizes  for  these  rare  classes.  At  the  other 
extreme,  rare  classes  may  be  assigned  low  priority,  and  an  equal  probability  design  is  a 
viable  option.  Rare  classes  will  not  be  sampled  in  large  numbers  in  such  designs.  An 
intermediate  emphasis  on  rare  land-cover  types  requires  a  compromise  design.  Two 
general  strategies  are  offered. 

The  first  strategy  is  to  treat  rare  classes  as  a  separate  sampling  problem.  For 
example,  a  general  purpose  sampling  design  for  accuracy  assessment  could  be  a  simple 
random,  systematic,  or  cluster  sample.  Because  such  designs  will  result  in  only  a  few 
reference  samples  in  the  rare  classes,  the  general  design  is  supplemented  by  a  specialized 


487 


design  tailored  to  sample  rare  classes  with  high  probability.  This  two-step  design  must 
still  be  conducted  according  to  a  probability  sampling  protocol,  and  the  analysis  must 
take  into  account  that  those  pixels  selected  in  the  supplemental  design  may  have 
different  inclusion  probabilities  than  those  pixels  selected  by  the  original  design. 

Adaptive  cluster  sampling  (Thompson  1990)  is  tailored  to  sample  efficiently  rare, 
but  spatially  clustered  items.  An  example  illustrating  this  strategy  in  an  accuracy 
assessment  setting  is  as  follows.  Suppose  the  rare  land-cover  class  is  marsh,  and  assume 
that  a  5x5  block  of  pixels  is  adopted  as  the  psu.  A  simple  random  or  systematic  sample 
of  psus  is  selected.  For  those  psus  in  which  at  least  one  marsh  pixel  (according  to  the 
reference  classification)  is  found,  the  sampling  procedure  is  "adapted"  to  then  sample 
adjacent  5x5  blocks  surrounding  the  initial  sampled  psu.  If  one  or  more  of  these 
adjacent  psus  is  found  to  have  at  least  one  marsh  pixel,  the  adaptive  strategy  is 
continued.  The  process  stops  when  no  marsh  pixels  are  found  in  adjacent  psvs.  If  the 
rare  cover-type  is  spatially  clustered,  this  strategy  will  greatly  improve  the  probability 
of  sampling  reference  pixels  in  this  class  because  the  method  intensifies  sampling  effort 
in  those  areas  in  which  the  rare  class  is  found.  The  adaptive  design  satisfies  the 
necessary  criteria  of  a  probability  sample.  Special  estimators  of  the  accuracy  parameters 
need  to  be  employed  when  this  design  is  used  (see  Thompson  (1990)  for  the  basic 
theory).  Adaptive  cluster  sampling  may  be  an  effective  design  for  change  detection 
accuracy  assessment  because  while  change  pixels  may  be  rare,  they  are  likely  to  be 
spatially  clustered,  and  the  adaptive  design  may  more  cost-effectively  sample  such 
pixels. 

Double  sampling  is  another  design  alternative  that  may  be  used  to  enhance 
estimation  for  rare  cover  types.  In  double  sampling,  a  large  first-phase  sample  is 
selected  and  the  stratum  to  which  each  sampled  pixel  belongs  is  identified.  From  these 
first-phase  sample  units,  a  stratified  random  sample,  called  the  second-phase  sample,  is 
then  selected.  The  primary  advantage  of  this  method  is  that  only  the  first-phase  sample 
units,  not  all  N  pixels,  need  to  be  assigned  to  strata.  The  stratification  could  be  based 
on  the  land-cover  classes  as  identified  by  the  reference  data  instead  of  stratifying  on  the 
land-cover  classes  identified  by  the  map.  This  stratification  is  then  used  to  increase  the 
probability  of  sampling  rare  ground  classes.  To  implement  this  design,  a  large  first- 
phase  sample  is  selected  and  each  location  is  assigned  to  a  land-cover  class.  These 
assignments  would  not  necessarily  require  a  ground  visit  if  reasonably  accurate  stratum 
identifications  can  be  made  using  available  maps,  aerial  photographs  or  videography. 
The  second-phase  sample  is  then  a  stratified  random  subsample  of  the  first-phase 
sample,  and  the  "true"  classification  of  these  second-phase  sample  sites  is  identified. 

To  illustrate  how  this  approach  focuses  sampling  effort  on  rare  categories,  suppose  a 
particular  class,  say  marsh,  represents  only  0.1%  of  the  land  area,  and  that  commission 
error  for  marsh  is  high.  If  the  stratification  is  based  on  land-cover  as  identified  by  the 
map,  it  is  likely  that  only  a  few  true  marsh  sites  will  appear  in  the  reference  sample.  By 
using  double  sampling  and  stratifying  on  the  ground  classification,  the  probability  of 
sampling  marsh  reference  pixels  could  be  increased  dramatically  by  intensifying 
sampling  effort  in  the  marsh  stratum.  Increasing  the  sample  size  in  this  manner  will 
improve  precision  of  producer's  accuracy  estimates  for  the  marsh  class. 

The  second  strategy  for  sampling  rare  classes  is  to  adapt  the  general  purpose 
sampling  design  itself  to  emphasize  more  strongly  rare  classes.  A  stratified  design  is  a 
typical  approach  for  such  a  one-step  strategy.  However,  if  the  general  design  is  chosen 
to  emphasize  sampling jrare  classes,  some  precision  will  be  lost  for  other  estimates  of 
map  accuracy  such  as  Pc  and  k  because  of  this  allocation  of  sampling  resources.  This 


488 


precision  trade-off  is  characteristic  of  any  sampling  design  choice,  and  the  properties  of 
the  strategy  selected  should  reflect  the  objectives  specified  for  the  accuracy  assessment. 

ENHANCING  PRECISION  VIA  ANALYSIS 

The  second  component  of  a  cost-effective  sampling  strategy  is  the  analysis  of  the 
data,  provided  by  the  estimators  of  map  accuracy  parameters.  Three  techniques  for 
improving  the  precision  of  map  accuracy  estimates  at  the  analysis  stage  are  post- 
stratification,  regression  estimation,  and  incorporating  "found"  data  into  an  accuracy 
assessment.  All  three  approaches  use  auxiliary  information  in  the  analysis  to  achieve 
this  precision  gain.  Poststratification  and  regression  estimation  can  be  applied  to 
practically  any  sampling  design.  Here  it  is  assumed  that  SRS  is  employed  because  an 
advantage  of  these  analysis  techniques  is  that  they  can  improve  precision  for  simple, 
easily  implemented  sampling  designs. 

Poststratification 

Poststratification  is  an  estimation  technique,  not  a  sampling  design,  requiring 
auxiliary  information  similar  to  that  required  for  stratified  sampling.  Suppose  n  sample 
pixels  are  obtained  via  SRS.  Let  Nfc_|_  and  nfc+  denote  the  number  of  pixels  classified  as 
cover  type  k  in  the  population  and  sample,  respectively.  Both  Nfc+  and  nk+  are  known 
once  the  map  is  completed.  Poststratification  incorporates  the  known  totals  Nfc+  into 
the  estimator  by  analyzing  the  SRS  data  as  a  stratified  random  sample  in  which  nfc+ 
pixels  have  been  selected  from  the  Nj._|_  pixels  available  in  that  stratum  (cover  type  k  as 
identified  on  the  map).  Consider  the  estimator  of  Pc.  If  nkk  is  the  number  of  sample 
pixels  correctly  classified  in  cover  type  k,  the  usual  SRS  estimator  for  Pc  is  ^n^/n, 
where  summation  is  over  the  q  cover  types.  For  SRS,  each  pixel  is  weighted  equally, 
the  weights  being  N/n.  For  the  poststratified  estimator,  the  weights  are  dependent  on 
the  identified  strata,  the  weight  for  stratum  k  being  Nfc+/nfc+.  Then  the  poststratified 
estimator  of  Pc  is  (1/N)  53  i=i(^  k+/nk+)nkk'  Poststratified  estimators  for  other 
parameters  are  constructed  in  essentially  the  same  manner,  replacing  the  weight  N/n 
used  in  the  usual  SRS  formula  by  N^+/nfc+,  and  then  summing  over  the  q  strata.  Card 
(1982)  presents  poststratified  estimators  and  variance  estimators  for  some  accuracy 
parameters,  and  Stehman  (1995)  shows  the  poststratified  estimators  of  Pc  and  K.  The 
precision  achieved  by  poststratification  is  approximately  that  of  a  stratified  sample  with 
proportional  allocation  (Cochran  1977,  p.  134),  so  poststratification  will  usually  result  in 
some  gain  in  precision  over  the  usual  SRS  estimators.  Based  on  a  small  simulation 
study  (Stehman  1996c),  the  gain  in  precision  from  poststratification  can  be  expected  to 
be  around  5%  (in  terms  of  standard  error)  for  estimating  Pc  and  /c,  and  as  much  as  15- 
30%  for  estimating  producer's  accuracy. 

A  disadvantage  of  poststratification  relative  to  a  stratified  design  is  that  the  sample 
size  within  each  stratum,  nfc_|_,  is  under  control  with  a  stratified  design,  but  not  with 
SRS  followed  by  poststratification.  However,  poststratification  can  be  applied  to  any 
classification  scheme.  For  example,  if  a  user  wishes  to  collapse  land-cover  classes  or  use 
an  entirely  different  classification  scheme,  poststratification  can  still  be  applied. 
Poststratification  is  flexible  because  it  can  be  adapted  to  any  identified  subpopulation, 
whereas  the  usual  stratified  sampling  design  is  advantageous  only  for  those 
subpopulations  identified  as  strata  prior  to  sampling.  Poststratification  and  stratified 
sampling  both  require  N^+,  but  these  are  available  once  the  land-cover  map  is 
completed.  Because  poststratification  does  not  require  this  information  to  obtain  the 
sample,  no  time  delay  between  the  imagery  and  the  ground  sampling  need  occur. 


489 


Regression  Estimator 


In  the  regression  estimator  approach,  any  auxiliary  information  available  that  can 
provide  a  land-cover  classification  of  a  pixel  may  be  used.  Primary  sources  of  auxiliary 
data  include  aerial  photography  or  videography,  or  even  another  land-cover  map 
obtained  via  remote  sensing  and  using  a  coarser  scale  of  resolution  (e.g.,  AVHRR).  This 
auxiliary  information  is  separate  from  the  imagery  data  used  to  construct  the  target 
land-cover  map  being  assessed.  A  sample  of  ground  reference  locations  is  still  required, 
and  these  reference  data  are  combined  with  the  auxiliary  data  via  a  regression  estimator 
to  estimate  Pc.  For  SRS  of  n  pixels,  the  sample  proportion  of  pixels  in  which  the 
reference  and  map  classifications  match  is  py.  If  no  auxiliary  data  are  available,  py  is 
the  usual  estimator  of  Pc.  The  sample  proportion  of  pixels  in  which  the  map  classi- 
fication matches  the  classification  obtained  by  the  auxiliary  data  is  px,  and  p  is  the 
sample  proportion  of  pixels  in  which  both  the  reference  and  auxiliary  data  classifications 
match  the  map  classification.  The  regression  estimator  of  Pc  is 

Pr.g  =  Py+gi(1_gi)(Px-Px).  [1] 

where  Px  is  the  proportion  of  pixels  in  the  entire  map  in  which  the  auxiliary  data  and 
map  classifications  match.  The  estimated  variance  of  preg  is 


var(pre^)  =  -n^ry 


-  \2 


1-n/N^..      _      (P*y  -  P*Py) 


[2] 


The  regression  estimator  results  in  some  gain  in  precision  over  the  usual  SRS 
estimator,  no  matter  how  poor  the  auxiliary  data  classifications,  but  the  approach  is  not 
worth  implementing  unless  a  meaningful  gain  in  precision  is  achieved.  The  more 
accurate  the  classifications  from  the  auxiliary  data,  the  greater  the  gain  in  precision 
achieved  by  the  regression  estimator  relative  to  the  usual  SRS  estimator. 

The  usual  regression  estimator  approach  requires  auxiliary  information  for  the  entire 
map  region.  In  practice,  it  is  more  feasible  to  use  a  double  sampling  approach  in  which 
the  auxiliary  data  are  collected  for  a  first-phase  sample,  and  ground  visits  are  made  to  a 
subsample  of  the  first-phase  sample.  The  double  sampling  regression  estimator  incor- 
porates both  the  first-  and  second-phase  sample  data.  Stehman  (1996a)  demonstrated 
that  the  precision  of  the  regression  estimator  combined  with  double  sampling  was  nearly 
the  same  as  the  strategy  employing  the  regression  estimator  with  SRS.  The  double 
sampling  strategy  is  more  cost-effective  because  it  requires  obtaining  the  auxiliary  data 
only  on  the  first-phase  sample,  not  the  entire  target  region. 


Found  Data 

Additional  potential  ground  reference  data  may  be  available  via  purposeful, 
haphazard,  convenience,  or  other  non-probability  sampling  methods.  For  example, 
suppose  an  organization  has  land-cover  classifications  for  various  special  interest  sites 
which  were  visited  "to  see  what  was  there."  Assuming  the  land-cover  classifications 
made  at  these  sites  are  correct  and  consistent  with  the  classification  scheme  employed 
for  the  map  being  evaluated,  is  it  possible  to  use  these  data  in  the  accuracy  assessment? 
Such  data  in  a  sense  represent  "free"  reference  samples  because  they  already  exist.  But 
incorporating  them  into  the  accuracy  assessment  in  a  statistically  valid  manner  is  not 
simple  because  it  is  difficult  to  generalize  from  these  data  to  the  population  at  large. 
That  is,  what  population  do  these  "found"  sites  represent?  To  illustrate  the  nature  of 
the  difficulty,  if  reference  samples  were  obtained  only  from  areas  within,  say,  250  meters 


490 


of  a  road,  the  sample  data  can  be  generalized  only  to  a  population  of  the  area  in  the 
study  region  within  250  meters  of  a  road.  Generalizing  beyond  that  area  is  not 
supported  by  statistical  inferences,  and  must  rest  on  non-statistical  arguments  regarding 
what  these  sites  represent.  The  same  can  be  said  of  "found"  sites.  Because  these  sites 
were  not  selected  by  probability  sampling  methods,  the  population  represented  by  these 
sites  is  unknown.  The  probability  sampling  protocol  must  be  replaced  by  assumptions 
concerning  the  population  represented  by  the  found  data.  Overton  (1990)  and  Overton 
et  al.  (1993)  describe  some  statistical  procedures  for  using  found  data  that  apply  to  this 
problem.  However,  the  amount  of  work  (cost)  needed  to  use  these  data  in  a  statistically 
valid  manner  appears  high,  and  unless  these  data  are  abundant  or  exceedingly  valuable, 
such  as  for  a  very  rare  cover-type,  it  is  doubtful  that  such  data  can  contribute 
significantly  to  reducing  the  costs  of  a  statistically  rigorous  accuracy  assessment. 
Incorporating  existing  data  is  much  more  feasible  if  a  valid  sampling  design  was  used  to 
obtain  the  data. 

GENERAL  RECOMMENDATIONS 

Some  general  recommendations  for  sampling  in  large-scale  accuracy  assessment 
projects  are  proposed.  As  with  any  general  recommendation,  it  is  easy  to  think  of 
situations  in  which  exceptions  would  obviously  be  needed.  Recommendations  should 
also  evolve  over  time  as  better  methods  and  new  insights  are  gained.  With  those 
caveats  in  mind,  the  following  suggestions  are  proposed.  For  accuracy  assessment  of 
large-area  land-cover  maps,  stratification  by  a  few  large  geographic  regions  is  helpful  to 
ensure  a  regionally  representative  sample,  to  provide  adequate  sample  sizes  in  these 
regions  for  reporting  region  specific  accuracy  statistics,  and  to  spread  the  workload  more 
evenly.  As  an  example  of  the  scale  of  the  recommended  stratification,  a  state  such  as 
North  Carolina  could  be  stratified  into  three  physiographic  regions,  the  coastal  plain, 
Piedmont,  and  western  upland  areas.  Bauer  et  al.  (1994)  provide  another  example,  as 
they  partitioned  their  study  region  (Minnesota)  into  eight  physiographic  regions  that 
could  be  used  as  strata  in  an  accuracy  assessment.  States  themselves  are 
administratively  convenient  strata  for  a  land-cover  map  spanning  several  states. 

Within  each  geographic  stratum,  a  simple  design  such  as  simple  random  or 
systematic  sampling  is  recommended  as  the  general  design.  Implementing  a  simple 
design  creates  the  opportunity  to  employ  precision  enhancing  analysis  techniques  such  as 
poststratification  and  regression  estimation.  Cluster  sampling  has  been  demonstrated  to 
be  cost-effective  (Moisen  et  al.  1994),  and  the  combination  of  cluster  sampling  with 
stratification  into  road  and  non-road  areas  (Edwards  et  al.  1996)  has  considerable 
practical  appeal.  Simple  designs  are  easier  to  implement  in  the  field  thus  increasing  the 
likelihood  that  the  design  will  be  implemented  correctly.  This  general  design  strategy  is 
adaptable  to  a  variety  of  analyses  and  classification  systems,  and  will  accommodate  the 
multiple  general  uses  and  objectives  present  in  a  large-area  mapping  project. 

To  accomplish  more  specialized  objectives,  a  design  tailored  to  these  objectives  can 
supplement  the  general  purpose  sampling  design.  For  example,  if  sampling  certain  rare 
classes  is  an  important  objective,  the  data  from  the  general  design  can  be  supplemented 
by  an  additional  simple  random  sample  from  each  rare  class.  Users  must  be  careful  to 
incorporate  proper  weights  in  the  analysis  when  combining  data  collected  by  these  two 
different  designs.  The  double  sampling  and  adaptive  cluster  sampling  designs  described 
earlier  may  also  be  used  to  supplement  a  simpler,  general  purpose  design. 

Accuracy  assessment  of  large-area  thematic  maps  creates  several  challenging 
statistical  problems.    These  problems,  however,  are  not  insurmountable.    The  classical 


491 


finite  sampling  approach  has  been  in  use  for  over  50  years  and  has  been  applied  to 
sampling  large,  spatially  disperse,  and  hard  to  measure  populations.  National  samples 
obtaining  labor,  economic,  and  health  statistics  are  applied  to  populations  over  large 
geographic  areas,  and  these  programs  face  a  demanding  set  of  objectives  requiring 
estimates  for  a  variety  of  parameters  at  various  spatial  scales.  These  problems  are 
characteristic  of  large-area  accuracy  assessment  efforts.  Classical  finite  sampling  theory 
and  methods  provide  a  rich  toolbox  from  which  to  choose  cost-effective  sampling 
strategies,  and  adapting  these  strategies  to  accuracy  assessment  of  large-area  thematic 
maps  is  a  viable  approach  to  pursue. 

ACKNOWLEDGMENTS 

I  thank  Ray  Czaplewski  for  his  review  and  helpful  suggestions.  This  work  has  been 
supported  by  cooperative  agreement  CR821782  between  the  U.S.  EPA  and  SUNY-ESF. 

REFERENCES 

Bauer,  M.E.  et  al.  1994.  Satellite  inventory  of  Minnesota  forest  resources.  Photogram. 

Eng.  &  Remote  Sensing  60:  287-298. 
Card,  D.H.  1982.  Using  known  map  category  marginal  frequencies  to  improve  estimates 

of  thematic  map  accuracy.  Photogram.  Eng.  &  Remote  Sensing  48:  431-439. 
Cochran,  W.G.  1977.  Sampling  Techniques  (3rd  ed).  Wiley:  New  York. 
Czaplewski,   R.L.   1994.   Variance  approximations  for  assessments  of  classification 

accuracy.    Res.  Pap.  RM-316.  Fort  Collins,  CO:  U.S.  Department  of  Agriculture, 

Forest  Service,  Rocky  Mountain  Forest  Range  and  Experiment  Station.  29p. 
Edwards,  T.C.,  Jr.,  Moisen,  G.G.,  and  Cutler,  D.R.  1996.  Assessing  map  accuracy  in  an 

ecoregion-scale  cover-map  (in  review). 
Janssen,  L.L.F.,  and  van  der  Wei,  F.J.M.  1994.  Accuracy  assessment  of  satellite  derived 

land-cover  data:  A  review.  Photogram.  Eng.  &  Remote  Sensing  60:  419-426. 
Moisen,  G.G.,  Edwards,  Jr.,  T.C.,  and  Cutler,  D.R.  1994.  Spatial  sampling  to  assess 

classification  accuracy  of  remotely  sensed  data.  In  Environmental  Information 

Management  and  Analysis:  Ecosystem  to  Global  Scales,  W.K.  Michener,  J.W. 

Brunt,  and  S.G.  Stafford  (eds).  New  York:  Taylor  and  Francis. 
Overton,  W.S.  1990.  A  strategy  for  use  of  found  samples  in  a  rigorous  monitoring 

design.  Tech.  Rep.  139,  Dept.  of  Statistics,  Oregon  State  University,  Corvallis,  OR. 
Overton,  J.McC,  Young,  T.C.,  and  Overton,  W.S.  1993.  Using  'found'  data  to  augment 

a  probability  sample:  procedure  and  case  study.  Envir.  Monit.  and  Assmt.  26:  65-83. 
Stehman,  S.V.  1995.  Thematic  map  accuracy  assessment  from  the  perspective  of  finite 

population  sampling,  Inter.  J.  of  Remote  Sensing  16:  589-593. 
Stehman,  S.V.  1996a.  Use  of  auxiliary  data  to  improve  the  precision  of  estimators  of 

thematic  map  accuracy  (Remote  Sensing  of  Environment,  in  review). 
Stehman,  S.V.  1996b.  Estimating  standard  errors  of  accuracy  assessment  statistics  under 

cluster  sampling  (Remote  Sensing  of  Environment,  in  review). 
Stehman,  S.V.  1996c.  Sampling  design  and  analysis  issues  for  thematic  map  accuracy 

assessment.  ASPRS  and  ACSM  Annual  Proceedings  (to  appear). 
Thompson,  S.K.  1990.  Adaptive  cluster  sampling.  J.  Amer.  Stat.  Assoc.  85:  1050-1059. 

BIOGRAPHICAL  SKETCH 

Stephen  Stehman  is  an  Associate  Professor  at  SUNY-ESF.  He  has  a  B.S.  in  Biology 
from  Penn  State,  an  M.S.  in  Statistics  from  Oregon  State,  and  a  Ph.D.  in  Statistics 
from  Cornell.  He  provides  statistical  consulting  for  faculty  and  graduate  students  at 
ESF  and  teaches  courses  in  sampling,  experimental  design,  and  multivariate  statistics. 


492 


Spatial  and  Probabilistic 
Classification  of  Forest  Structures 
Using  Landsat  TM  Data 

Jeffrey  L.  Moffett1  and  Julian  Besag2 

Abstract.  Satellite  sensors  record  upwelling  radiant  flux  from  the  Earth's 
surface.  Classifying  forest  structures  from  these  measurements  is  a 
statistical  inference  problem,  assuming  the  classes  are  defined  at  scales 
corresponding  to  the  spectral  and  spatial  resolution  of  the  data.  Bayesian 
image  analysis  methods  incorporate  both  spectral  and  spatial  (contextual) 
information  to  improve  classification  accuracy  over  spectral  classifiers. 
Stochastic  simulation  techniques  associated  with  Bayesian  models  provide 
a  mechanism  for  estimating  class  probability  vectors  for  each  image  pixel. 
This  study  evaluates  a  Markov  chain  Monte  Carlo  (MCMC)  technique, 
the  Gibbs  sampler,  for  forest  classification.  The  Gibbs  sampler  runs  as  a 
time-homogeneous  Markov  chain.  Class  estimates  are  obtained  by  seeking 
the  marginal  posterior  modes  (MPM)  of  the  limit  distribution.  The  image 
is  topographically  normalized  and  training  samples  are  collected  from 
randomly  distributed  pixels.  Preliminary  results  using  a  simple  model 
show  an  improvement  over  maximum  likelihood.  A  hierarchical  model 
designed  to  incorporate  the  terrain  effects  and  the  system  point  spread 
function  (PSF)  is  currently  under  development. 


INTRODUCTION 

Inferring  forest  stand  types  from  image  data  is  a  classification  process  that 
reduces  raw  data  to  useful  information.  Satellite  remote  sensing  can  be  useful  as 
a  means  of  stratifying  land  cover  in  the  first  stage  of  a  multi-stage  sampling 
scheme  for  forest  inventory.  Classified  maps  also  provide  necessary  information 
for  landscape  planning  such  as  the  distribution  of  stand  structure  classes  by  area 
across  a  watershed.  Probabilistic  contextual  classifiers  are  well  suited  to  modelling 
the  uncertainty  inherent  in  the  process. 

BAYESIAN  IMAGE  ANALYSIS  METHODOLOGY 

For  a  given  scale,  the  desired  classification  can  be  modelled  as  a  discrete 
Markov  random  field  (MRF)  x  with  an  associated  probability  distribution  tt(x). 


Graduate  student.  Silviculture  Laboratory,  University  of  Washington,  Seattle,  WA  (jmoffett@silvae.cfr.washington.edu). 

t 

Professor,  Department  of  Statistics,  University  of  Washington,  Seattle,  WA. 


493 


This  probability  distribution  can  be  combined  with  a  likelihood  density  for  the 
data  y  given  x.  The  combination  gives  a  posterior  probability  distribution  for  the 
random  field.  The  posterior  distribution  is  thus  a  conditional  probability 
distribution  for  the  classification  given  the  image  data.  Statistical  inferences  about 
x  are  then  based  on  the  posterior  distribution  ir(x\y). 

There  are  several  advantages  of  this  approach.  In  addition  to  classifying  each 
pixel  on  the  basis  of  the  spectral  information  contained  in  the  image  data,  other 
a  priori  information  can  be  specifically  modelled  by  the  prior  distribution.  This 
information  pertains  to  the  local  characteristics  of  the  classification,  e.g.  having 
adjacent  pixels  belonging  to  the  same  class,  having  objects  with  closed  boundaries, 
or  preventing  two  particular  classes  from  being  adjacent.  Unlike  post-classification 
smoothing  algorithms,  such  as  majority  filters,  Bayesian  classification  techniques 
incorporate  spatial  information  without  losing  fidelity  to  the  data. 

Observed  data 

Consider  a  two-dimensional  rectangular  array  of  pixels  over  a  finite  region  S 
with  each  pixel  identified  by  an  index  i  =  1,2,  n.  For  each  pixel  in  S  a  remote 
sensor  records  a  measurement  of  radiant  flux  that  is  a  sample  distribution  of  the 
upwelling  radiance,  degraded  by  a  stochastic  process.  The  imperfect  observations 
form  an  image  y  =  {y,:  i  e  S}.  The  yf's  are  vector  valued  for  bands  1-5,7. 

Markov  random  fields 

Let  X  be  a  random  vector  that  classifies  y  over  region  S  and  x  =  {*,,•  i  e  S]  be 
any  realization  of  X.  Thus,  x  denotes  an  arbitrary  classification  of  S  and  X  = 
{Xj,...,Xn}.  Each  univariate  Xt  takes  values  among  c  unordered  classes.  The 
objective  of  classification  is  to  estimate  the  "true"  class  image  x*.  Bayesian  image 
analysis  assumes  x*  is  a  realization  of  a  locally  dependent  MRF  with  an  associated 
joint  probability  distribution  ir(x)  assigning  classes  to  S  (Besag,  1986).  In  forest 
type  mapping  stands  are  typically  larger  than  the  individual  pixels.  Exploiting  this 
spatial  continuity  of  the  surface  cover  at  a  particular  scale  relative  to  the  image 
resolution  provides  a  second  source  of  imperfect  information  about  x*. 

Unlike  time  dependent  variables  the  only  logical  conditioning  set  for  spatial 
variables  is  all  x  except  xt.  An  MRF  must  be  defined  with  respect  to  a  set  of 
neighbors,  written  as  xdi,  and  satisfy  the  positivity  condition:  tt(x)  >  0  for  all  x. 
Given  a  specified  neighborhood  system,  the  Markov  property  states  7r(jtf|jcf)  = 
rr{xi  \xdl),  where  x4  represents  all  pixels  other  than  pixel  i.  The  Markov  property 
allows  the  full  conditional  distribution  of  xt  to  be  specified  in  terms  of  its 
neighbors.  As  a  result  of  this  assumption,  ir(x)  is  a  Gibbs  distribution  with  respect 
to  the  assumed  neighborhood  (Geman  and  Geman,  1984).  The  MRF-Gibbs 
equivalence  provides  an  explicit  formula  for  the  jointly  distributed  random  field, 
in  terms  of  the  full  conditionals  or  Gibbs  potential  functions. 


494 


A  second-order  neighborhood  consists  of  eight  neighbors  in  the  horizontal, 
vertical,  and  diagonal  directions.  The  general  form  of  a  pairwise  interaction  MRF 
for  discrete  unordered  classes  can  be  written  as  follows  (Besag,  1986): 


tt(x)  oc  exp( 


\<k<l<c  Pkpkl  h 


\<k<c 


where  nk  is  the  number  of  pixels  in  class  k  and  nkl  is  the  number  of  distinct  class 
pairs  (k,l).  The  aks  and  the  j3w's  govern  the  percentage  of  pixels  in  each  class  and 
the  strength  of  interaction  between  different  classes;  for  (3kl  >  0  classes  k  and  /  are 
discouraged  from  being  neighbors. 

Likelihood  function 

The  recorded  spectral  radiance  influences  the  classification  through  the 
likelihood  function.  In  contrast  to  the  prior,  the  likelihood  is  ideally  determined 
by  knowledge  of  the  sensing  process  and  the  existence  of  training  data.  The 
likelihood  function  represents  the  joint  probability  density  as  a  conditional  density 
of  the  data  y  given  the  actual  scene  x: 


Written  in  this  form  the  image  data  is  assumed  to  be  conditionally  independent 
given  knowledge  of  the  classes,  or  class  conditionally  independent.  While  this 
assumption  expedites  an  overview  of  the  theory,  and  may  be  reasonable  for  some 
applications,  it  is  not  a  valid  assumption  for  remotely  sensed  images. 

Posterior  distribution 

Bayes  theorem  can  be  used  to  construct  a  posterior  distribution  rr(x \y)  on  which 
inferences  about  x*  can  be  based.  Using  this  theorem,  the  posterior  distribution  for 
the  scene  x  given  the  data  y  can  be  written  as, 


The  computational  burden  presented  by  the  posterior  distribution  is  overcome 
by  the  fact  that  it  too  is  an  MRF  with  a  Gibbs  distribution.  This  allows  pixels  to 
be  evaluated  individually  based  on  their  full  conditional  distributions,  written  as 


L(y\x)  =1^/(3,.  I  jr.). 


ir(x\y)*L(y\x)ir(x). 


7T(  *.  |  X_tf  )  oc  /(  y.  |  Xi  )  7T(  X.  |  Xs  ). 


The  conditional  prior  distribution  follows  from  tt(x)  =  7r(jtf  |jc_ ^(jc.,). 


495 


Image  estimation 

Within  this  framework  the  final  classification  is  a  point  estimate  of  x*.  Due  to 
the  high  dimensionality  of  x  the  optimization  required  to  find  an  exact  solution  is 
infeasible  for  realistic  scenes.  Estimates  of  various  image  attributes  can  be 
obtained  using  stochastic  simulation.  The  marginal  posterior  modes  (MPM) 
estimate  will  minimize  the  expected  number  of  misclassifications  by  maximizing 
the  marginal  probability  with  which  each  pixel  is  classified.  MPM  has  been 
increasingly  adopted  in  recent  years  for  classification  problems  (Besag  et  aL, 
1991;  Green  et  ai,  1994).  MPM  estimates  of  x*  are  obtained  by  maximizing 
^(X  b)>  where  the  most  frequently  sampled  class  for  pixel  i  is  the  estimate  of 

This  estimate  can  be  computed  using  the  Gibbs  sampler  (Geman  and  Geman, 
1984)  which  runs  as  a  discrete-time  Markov  chain,  with  state  space  the  set  of  all 
possible  classifications  x  and  limit  distribution  ir{x\y)  (Besag,  1989).  For  a  single 
iteration,  a  new  xt  is  randomly  sampled  for  each  pixel  from  its  full  conditional 
distribution.  This  generates  a  sequence  of  stochastically  dependent  classifications 
{x}.  An  initial  classification  is  required  to  start  the  algorithm  followed  by  a 
number  of  iterations  during  which  samples  are  not  collected.  This  burn-in  period 
will  minimize  the  influence  of  the  initial  classification.  This  procedure  also 
estimates  a  posterior  probability  distribution  for  each  pixel,  providing  spatially 
explicit  information  about  the  certainty  of  the  point  estimate  classification. 
Credible  intervals  can  also  be  obtained  for  measurements  such  as  class  areas.  The 
MPM  estimate  requires  simulating  a  predetermined  number  of  realizations  of  the 
MRP  and  it  can  be  difficult  to  determine  the  number  of  realizations  required  to 
assure  convergence.  Parameter  estimation  can  also  be  problematic. 

A  PRELIMINARY  MODEL 

Using  this  approach  a  project  was  undertaken  to  classify  TM  data  for  the  Lizard 
Lake  watershed  near  Clallam  Bay,  Washington.  The  area  is  covered  by  second 
growth  forests.  Six  classes  were  selected  to  represent  general  forest  structures  and 
land  cover:  clearcut  (CC);  stand  initiation  (SI);  stem  exclusion  (SE);  older  forest 
(OF);  hardwood  (HD);  and  deep  water  (W).  Class  means  and  covariances  were 
estimated  from  randomly  selected  pixels  within  training  sites  to  avoid  biasing  the 
estimates.  As  this  is  an  area  of  rugged  terrain,  the  backwards  radiance  correction 
transformation  (Colby,  1991)  was  used  to  topographically  normalize  the  image. 

The  model 

A  simple  model  was  constructed  by  first  adopting  the  conventional  maximum 
likelihood  function  (ML),  which  assumes  the  data  to  be  class  conditionally 
independent  and  normally  distributed.  A  discrete  pairwise  difference  prior  was 
chosen  to  model  the  tendency  of  adjacent  pixels  to  be  of  the  same  class.  The  prior 
distribution  is  written  as  follows, 


496 


7T  (  x  )  :    7T  (*.  )  oc  e^n«(c), 


where  fi  is  a  strength  of  interaction  parameter,  and  rc/c)  counts  the  number  of 
neighbors  (8  nearest)  of  pixel  /  in  class  c  for  c  -  (0,...,5). 

Results 

The  Gibbs  sampler  was  written  in  C.  Samples  were  collected  for  3000  iterations 
after  a  burn-in  of  1000  iterations;^  was  fixed  at  0.5.  Figure  1  shows  the  "ground 
truth"  classification,  which  does  not  indicate  the  presence  of  hardwood  patches 
known  to  be  on  this  landscape.  Figure  2  shows  the  ML  estimate  from  data  that 
was  not  normalized.  The  topographic  shade  causes  some  SE  forests  to  be 
classified  as  OF.  Figure  3  shows  the  MPM  point  estimate  from  normalized  data. 
Figure  3  also  shows  a  single  pixel  at  the  center  of  a  small  lake  where  the  water 
is  deepest  (lower  left  corner).  The  influence  of  the  likelihood  preserves  the  single 
deep  water  pixel,  which  would  be  lost  using  a  majority  filter.  Figure  4  shows  a 
posterior  probability  classification  of  SE  forest.  Note  that  the  edges  of  the  younger 
forest  types  and  older  forests  have  a  modest  probability  of  being  stem  exclusion. 

A  HIERARCHICAL  MODEL 

The  likelihood  in  the  above  model  is  problematic  and  the  prior  is  the  most 
basic.  Sources  of  variation  such  as  atmospheric  scattering  and  topography 
complicate  the  classification  process.  As  an  alternative  to  data  preprocessing,  the 
Bayesian  approach  allows  separate  models  for  each  source  of  variation  to  be 
linked  through  successive  conditioning  of  the  form  7t(jc|z;),  7t(z;|z2),  Tr{zk\y) 
(Smith  and  Roberts,  1993).  Each  source  of  variation  is  modelled  by  a  conditional 
probability  distribution  as  a  stage  within  a  hierarchical  framework.  If  a  second- 
order  neighborhood  is  assumed,  triple  and  quadruple  pixel  interaction  can  be 
modelled  (Besag,  1974).  Triples  are  useful  for  modelling  convex  stand  boundaries 
and  quadruples  for  modelling  interior  regions.  Recent  research  suggests  that  for 
restoration  problems  a  crude  prior  will  suffice,  but  that  estimates  of  attributes  such 
as  class  area  and  perimeter  are  sensitive  to  the  specification  of  the  prior 
distribution  (Tjelmeland  and  Besag,  1996). 

Topographic  effect  model 

Estimates  of  slope  and  aspect  are  generally  calculated  from  digital  elevation 
models  (DEMs),  but  are  subject  to  several  sources  of  error.  Given  the  likely  errors 
and  little  knowledge  of  the  surface  photometric  function  it  is  therefore  logical  to 
model  the  effect  of  topography  with  some  degree  of  uncertainty.  This  can  be 
specified  by  first  defining  z  to  be  the  measured  pixel  image  incorporating  the 
effects  of  topography.  Assuming  z  to  be  conditionally  independent  given  x,  and 
each  Zj  to  be  normally  distributed  gives  the  following  model, 


497 


Figure  1.  "Ground  truth."  Note 
missing  patches  of  hardwoods. 


Figure  2.  ML  classification  without 
topographic  normalization  showing  a 
greater  portion  of  older  forest. 


Figure  3.  MPM  point  estimate,  /3=0.5. 


H  w     Si  si 

□    HD     ill  SE 
CZ3    CC     ■  OF 
Legend  for  Figs.  1,2,3. 


Figure  4.  Posterior  probability  of 
SE  forest,  /3=0.5.  Linear  stretch  with 
brighter  shades  indicating  higher 
probability. 


498 


cosA7  coske 


cose 


)  M,/), 


where  cos  i  is  the  cosine  of  the  angle  of  incidence,  cos  e  is  the  cosine  of  the 
slope,  6  is  the  variance,  /jx  are  the  means,  and  k  is  the  Minnaert  constant.  This 
model  is  derived  from  the  normalization  transformation  of  Colby  (1991).  The 
assumption  of  normality  in  this  model  is  arbitrary,  requiring  further  investigation. 

System  PSF  effects  model 

After  reflecting  off  of  a  topographic  surface  the  upwelling  radiant  flux  is 
scattered  by  the  atmosphere  and  recorded  as  a  function  of  the  sensor  PSF  (Duggin 
and  Robinove,  1990).  If  the  data  y  are  assumed  normally  distributed  and 
conditionally  independent  given  z,  then  this  process  can  be  modelled  as  follows, 


where  h  is  the  system  PSF  detecting  radiation  from  j  pixels  in  the  proximity  of 
pixel  i,  and  </>  is  the  variance.  Note  also  that  h  will  be  a  function  of  wavelength. 

Posterior  distribution 

Bayes  theorem  allows  these  models  to  be  combined  and  gives  the  following 
joint  posterior  distribution, 


Again  an  MPM  estimate  can  be  obtained  by  sampling  from  the  full  conditionals. 
However  the  added  complexity  of  this  model  may  require  the  use  of  a  Metropolis 
algorithm.  The  complexity  will  be  further  increased  if  prior  distributions  are 
specified  for  some  of  the  parameters  as  a  fully  Bayesian  approach  would  suggest. 


Appropriate  accuracy  assessment  methods  for  probabilistic  classifications  must 
still  be  developed.  Estimates  of  posterior  probabilities  will  be  difficult  to  verify. 
The  "ground  truth"  for  assessing  point  estimates  needs  to  be  at  the  same  scale 
and  resolution  as  the  image  data.  Sensitivity  analysis  must  be  conducted  to  assess 
the  stability  of  the  solutions.  Topographic  normalization  and  spatial  information 
can  improve  forest  classification.  Probabilistic  classifiers  provide  additional 
information  by  estimating  the  level  of  certainty  with  which  a  pixel  is  classified, 


tt(  y  |  z  )  a  exp(-^-  £  (y,     £  h..  z.)2), 


7T(  x,  z  |  y  )  <*  tt(  y  I  z  )  7r(  z  I  X  )  7T(  X  ). 


CONCLUSION 


499 


which  is  the  logical  result  of  an  inference  procedure. 

ACKNOWLEDGEMENTS 
Many  thanks  to  C.  Oliver,  G.  Klock,  J.  McCarter,  G.  Schreuder,  and  M.  Smith. 

REFERENCES 

Besag,  Julian.  1974.  Spatial  Interaction  and  the  Statistical  Analysis  of  Lattice 
Systems.  Journal  of  the  Royal  Statistical  Society,  Series  B.  36(2):  192-236. 

Besag,  Julian.  1986.  Statistical  Analysis  of  Dirty  Pictures.  Journal  of  the  Royal 
Statistical  Society,  Series  B.  48:259-302. 

Besag,  J.  1989.  Towards  Bayesian  Image  Analysis.  J.  ofAppl.  Stat.  16:395-407. 

Besag,  J.,  J.  York,  and  A.  Mollie.  1991.  Bayesian  Image  Restoration  with  Two 
Applications  in  Spatial  Statistics.  Ann.  of  the  Ins.  of  Stat.  Math.  43(1):  1-59. 

Colby,  Jeffrey  D.  1991.  Topographic  Normalization  in  Rugged  Terrain. 
Photo grammetric  Engineering  and  Remote  Sensing.  57(5):53 1-537. 

Duggin,  M.  and  C.  Robinove.  1990.  Assumptions  Implicit  in  Remote  Sensing 
Data  Acquisition  and  Analysis.  Int.  J.  of  Remote  Sensing.  11(10):  1669-1694. 

Geman,  S.  and  D.  Geman.  1984.  Stochastic  Relaxation,  Gibbs  Distributions, 
and  the  Bayesian  Restoration  of  Images.  IEEE  Trans.  RAMI.  6(6) :72 1-741. 

Green,  E.J.  et  al.  1994.  Construction  of  Thematic  Maps  from  Satellite  Imagery. 
In:  Bayesian  Statistics  5.  Proceedings  of  the  5th  Valencia  Int.  Meeting,  p.  75. 

Smith,  A.  and  G.  Roberts.  1993.  Bayesian  Computation  via  the  Gibbs  Sampler  and 
Related  Markov  Chain  Monte  Carlo  Methods.  J.  R.  Stat.  Soc,  B.  55(l):3-23. 

Tjelmeland,  Hakon  and  Julian  Besag.  1996.  Markov  Random  Fields  with 
Higher  Order  Interactions,  submitted  for  publication. 

BIOGRAPHICAL  SKETCH 

Jeffrey  Moffett  is  researching  MCMC  classification  of  RS  imagery  for  a  PhD 
dissertation  and  provides  GIS  support  for  the  Landscape  Management  Project. 

Julian  Besag  works  mainly  in  spatial  statistics  and  Bayesian  inference.  He  is  the 
Director  of  the  Center  for  Spatial  Statistics  at  the  University  of  Washington. 


500 


Total  Error  Estimation  in  a  Spatial 

Database  for  GIS 


Jose  Alberto  Quintanilha  and  Marcos  Rodrigues2 


Abstract.  -  The  aim  of  this  paper  is  to  present  a  methodology  to  estimate  the  errors 
in  digital  spatial  database  originated  from  the  digitizing  a  large  number  of  maps.  It 
is  undestood  that  the  desired  quality  of  that  digital  database  is  associated  to  a 
specific  foreseen  use.  Supposing  that  the  spatial  database  is  the  population  to  be 
sampled  and  the  objects  on  it  as  the  primary'  sampling  units,  two  strategies  are 
discussed:  the  spatial,  focused  on  areas  where  errors  are  likely  to  occur  and  the  non- 
spatial  sampling,  focused  on  the  selection  of  objects  according  to  a  database 
structure.  The  errors  types  are  organized  in  a  sequential  structure  upon  which  a 
multi-stage  sampling  strategy'  is  applied.  The  sampling  areas  are  selected  by  a 
specific  sampling  frame,  quadrats,  selected  according  to  spatial  variability  evaluated 
by  entropy  measures.  Supposing  that  the  amount  of  errors  in  the  spatial  units  have  a 
Poisson  distribution,  the  total  error  distribution  can  be  estimated,  allowing  thus  the 
inference  of  the  whole  population  quality. 


INTRODUCTION 

The  purpose  of  this  paper  is  to  present  a  methodology  to  estimate  the  total 
errors  in  a  spatial  database  for  GIS  applications.  The  population  to  be  sampled  is 
the  database  from  the  digitalization  of  a  large  amount  of  maps  in  paper  format. 

The  adopted  sampling  schema  envolve  spatial  and  non-spatial  aspects. The 
spatial  aspects  concern  the  selection  of  spatial  areas  were  errors  are  more  likely  to 
occur;  the  non-spatial  aspects  are  related  to  the  selection  of  spatial  objects  (points, 
lines  and  polygons)  structured  in  a  particular  data  model. 

The  sampling  area  frame,  quadrats,  are  selected  according  to  their  variability, 
measured  by  their  correspondent  entropy. 

A  sequence  of  non-independent  errors  in  database:  identification 
(completeness),  absolute  position,  relative  position  and  topology,  in  that  order  (for 
more  details,  see  Quintanilha,  1996),  was  used  under  a  multiple  stage  sampling 
procedure. 

The  total  errors  in  a  spatial  database  can  be  estimate  under  the  assumption  that 
the  frequencies  of  the  sequence  of  the  errors  follow  a  multivariate  Poisson 
distribution  (Johnson  and  Kotz,  1969). 


L'niversidade  de  Sao  Paulo  -  Brazil 
2  Universtdade  de  Sao  Paulo,  CARTA  Consultoha  -  Brazil 


501 


The  aim  is  to  obtain  an  estimate  for  each  type  of  error  examining  the  smallest 
number  of  maps.  A  methodology  to  analyse  the  frequency  of  each  type  of  errors 
(completeness,  absolute  position,  relative  position  and  topology)  in  a  spatial 
database  is  proposed. 

EVALUATION  OF  THE  ERRORS  IN  A  SPATIAL  DATABASE 

As  mentionned  before,  the  evaluation  of  the  sequence  of  errors  is  made  by  the 
following  steps: 

step  1 :  extract  a  sample  set  of  maps  (random  or  intentional)  from  the  different 

existent  sets  of  maps  (non-spatial  sampling); 
step  2:  extract  a  sample  of  maps  from  the  selected  set  of  maps  from  the  first 

step  (non-spatial  sampling); 
step  3 :  choose  areas  from  the  selected  maps  from  step  2  (spatial  sampling); 
step  4:  stratify  selected  areas  by  themes  or  coverage  (non-spatial 

stratification); 

step  5:  select  the  objects  from  the  database  by  themes  or  coverages  of  interest 
(non- spatial  sampling). 

In  steps  1,  2,  3  and  5,  Shannon's  entropy  concept  is  used  to  select  the  set  of 
maps,  the  maps  and  the  area  to  be  sampled  within  each  map  by  themes  and  by 
objects.  Considering  that  high  variability  implies  high  probability  to  observe  the 
errors  under  study,  one  selects  the  set  of  maps,  maps,  and  areas  with  higher 
entropy  value,  which  may  be  estimated  by  the  number  of  bits  of  the  correspondent 
datafile.  So,  sampling  spatial  database  was  reduced  to  locate  the  spatial  sampling 
units  using  information  theory. 

SPATIAL  SAMPLING 

Spatial  Sampling  Techniques 

We  aim  at  selecting  an  adequate  area  to  identify  and  count  errors  to  certify  the 
database  accuracy.  Therefore  all  the  spatial  units  and  spatial  sampling  frames  to  be 
used  are  area  type. 

The  most  common  sampling  frame  in  spatial  sampling  is  the  quadrat.  A  quadrat 
(Shaw  and  Wheeler,  1985)  is  the  result  of  a  superposition  of  a  regular  grid  on  the 
map  (or  area).  Errors  are  counted  within  the  quadrat. 

Kish  (1965),  affirms  that  quadrat  sampling  is  not  a  sampling  method,  but  a 
methodology  to  collect  information  in  a  spatial  position  selected  by  some 
traditional  sampling  scheme  (random,  sistematic,  stratified,  etc.). 

The  quadrat  sampling  is  recommended  when  the  units  of  the  population  are  not 
available  (either  access  to  them  is  difficult  or  the  associated  cost  is  very  high).  The 


502 


case  of  errors  evaluation  in  database  is  an  example  of  the  inviability  of  census  for 
reasons  of  time  and  cost. 

The  quadrats  sampling  can  be  used  in  two  different  purposes:  measure  the 
abundance  per  areal  unit  or  investigate  the  population  pattern  (Ripley,  1981).  In 
the  first  case,  one  may  suppose  that  the  errors  in  the  quadrat  follow  a  Poisson 
process  with  tax  X  (error  rate  i.e.,  number  of  errors  occurrence  per  area  unit).  The 
frequency  of  errors  in  a  quadrat  with  area  A  will  follow  a  Poisson  distribution 
(mean  and  variance  equal  to  XA).  In  the  second  case,  indexes  have  been  proposed 
that  are  based  on  counting  from  a  set  of  samples  of  quadrats  (Ripley,op.cit  ). 


1)  if  N(A)  is  the  observed  number  of  errors  in  area  A  and  Pn(A)  =  prob  {N(A) 
=  n},  3  X>0,  constant,  implies  that  for  any  quadrat  with  area  8A>0  (Diggle,  1979 
and  1981): 


where  o(5A)  denotes  magnitude  orders  of  real  numbers  and  vectors  sequences. 
2)  if  for  any  two  exclusive  and  congruent  regions,  A  and  B,  N(A)  e  N(B)  are 
independents.  Then: 

Pn(A)=  e("^A)  (XA)  n  /  n!  (n=0,l,...). 

Thus,  N(A)  follows  a  Poisson  distribution  with  mean  equal  to  XA. 

Determination  of  quadrats  size  or  number  of  quadrats 

The  determination  of  the  number  of  quadrats  are  related  to  a  cost  model,  which 
envolves  the  precision  of  the  estimates  and,  obviously,  the  determination  of  costs 
and  times.  For  a  certain  budget  or  duration  for  the  database  quality  investigation, 
the  spatial  objects  in  the  set  of  maps,  maps,  areas,  themes/coverage  can  be 
investigate.  In  others  words,  for  an  expect  value  for  the  error  of  the  estimatives, 
we  can  determine  an  ideal  size  of  the  sample  to  satisfy  this  precision. 

Certainly,  a  higher  number  of  small  quadrats  will  cover  the  area  more 
efficiently.  So,  the  chance  to  detect  some  type  of  error  will  be  improved  and  the 
correspondent  spatial  correlation  will  be  minimized.  However,  when  sampling 
units  types  are  areas  and  lines,  the  definition  of  the  quadrats  must  consider  the 
problem  of  the  size  of  these  objects  which  can  transcend  the  area  of  the  quadrat. 
One  example  is  the  segmentation  of  polygons  in  meaningless  lines. 


Formally: 


Pn(6A) 


1  -  AM  +  o(5A) 
X8A+  o(8A) 
o(5A) 


(n=0) 
(n=l) 
(n=2,3,...) 


503 


Determination  of  position  of  the  quadrats 

For  a  certain  area  represented  in  a  quadrat  and  for  a  certain  representation 
model,  raster  or  vector,  the  size  of  datafile  which  represents  that  quadrat  is 
variable,  proportional  to  the  number  c  of  differents  objects  inside  it.  For  each 
selected  area,  pj  is  the  probability  of  occurrence  of  the  ith  object,  which  is 
estimated  by  their  frequency.  The  size  of  the  datafile  related  to  a  quadrat  is 
proportional  to  the  number  c  too,  i.e.: 

c 

Zpi  =  1-00 
i=l 

Considering  t[  as  the  size  of  a  part  of  a  quadrat  datafile  related  to  object  i,  the 
expected  mean  size  of  the  quadrat  datafile  is  given  by: 

c 

E  (t)  =  Z  pj  tj 
i=l 

The  expected  size  of  the  quadrat  datafile  E  (t)  is  proportional  to  the  sum  of  the 
number  of  different  objects  within  it. 

Similarly,  the  entropy  of  the  quadrat  is  given  by: 

c 

H  =  -  Z  Pi  log2  pi  bits/quadrat . 
i=l 

So,  the  sampling  procedure  must  be  restricted  to  areas  where  the  entropy  is 
higher,  which  corresponds  to  the  areas  with  higher  number  of  bits.  The  maps  and 
sets  of  maps  will  be  selected  similarly. 

The  themes/coverage,  will  be  used  like  a  secondary  variable  in  the  objects 
stratitification:  objects  associated  with  each  theme/coverage  in  areas  with  the 
higher  quadrat  datafile  will  be  in  the  sample. 

Use  of  the  thematic  stratification 

The  option  for  a  stratified  sampling  by  themes/coverages  is  justified  since  the 
simple  random  sampling,  relates  to  the  global  distribution  of  the  population  (i.e., 
the  errors),  which  implies  that  some  type  of  errors  may  not  be  sampled  or  may  be 
undersampled.  This  could  be  avoided  with  a  larger  sample  but  then,  there  would 
implications  on  the  time  and  associated  costs. 


504 


TEST  FOR  THE  TOTAL  ERRORS  AMOUNT 


As  mentionned,  we  use  the  sequence  present  above  and  consider  each  type  of 
error  (identification,  absolute  position,  relative  position,  and  topology)  as  a 
Bernouille  experiment.  That  is,  there  are  two  possible  results  for  each  type  of 
error:  it  exist  or  not. 

The  assumptions  are: 

•  the  vector  Y  =  (Yj,  Y^,  Yr,  Y  j)'  is  the  vector  of  the  number  of 
errors  each  type  errors  in  n  quadrats,  where: 

Yj  is  the  number  of  errors  of  identification, 

Y\  is  the  number  of  errors  of  absolute  position  which  was  not 

generated  by  errors  of  identification, 
Yr  is  the  number  of  relative  position  which  was  not  generated 

by  the  absolute  position  errors, 
Yj  is  the  number  of  errors  of  topology  which  was  not  generated 

by  the  relative  position  errors; 

•  each  type  of  errors  follows  a  Poisson  distribution; 

•  the  four  kinds  of  errors  are  correlated  (identification  errors  implies 
absolute  position  errors  which  implies  relative  position  errors  which  implies 
in  errors  in  topology). 

So,  the  vector  Y  will  follow  a  Poisson  multivariate  distribution  with  the 

parameters:  X*  =  (X\,  X/±,  Xr,  Xj^  where  X{  is  the  tax  of  error  type  i,  where  i=  I 
(identification),  A  (absolute  position),  R  (relative  position),  T  (topology).  Details 
of  this  distribution  can  be  found  in  Johnson  and  Kotz  (1969);  Jensen  (1985)  e  Ho 
(1995). 

PROPOSED  METHODOLOGY 

Consider  a  situation  when  a  number  of  maps  with  the  same  cartographic 
characteristics  (scale,  datum,  coordinate  system,  generalization  conditions, 
legends,  etc.)  is  necessary  to  cover  a  whole  area.  Over  that  maps,  one  overlays  a 
regular  grid  which  create  a  similar  number  of  equal  quadrats  in  each  map.  The 
overall  procedure  is: 

•  create  a  homogeneous  sets  of  maps, 

•  select  a  sample  of  set  of  maps  according  of  the  entropy  criteria, 

•  for  each  selected  set  of  maps,  extract  a  sample  of  maps  according  the  same 
criteria  (entropy), 

•  for  each  map  in  the  sample,  locate  the  quadrats  with  the  higher  entropy, 

•  for  each  theme/coverage  in  each  selected  quadrat,  compute: 

1 .        the  number  of  identification  errors, 


505 


2.  the  number  of  absolute  position  errors  not  generated  by  the 
identification  errors; 

3 .  the  number  of  relative  position  errors  not  generated  by  the  previous 
errors; 

4.  the  number  of  topology  errors  not  generated  by  the  previous  errors. 
The  event  error  can  be  represented  by: 

error  =  Iu(~InA)  u  (~I  n  ~A  u  R)  u  (~I  n4n~Rn  T). 
Because  they  are  exclusive  events,  the  error  occurency  probability  can  be 
estimated  by: 

p(error)=  p(I)  or  p(~I  and  A)  or  p(~I  and  ~A  and  R)  or  p(~I  and  ~A  and 

~R  and  T)  = 

=  p(I)  +  p(~I  n  A)  +  p(~I  n  ~A  n  R)  +  p(~I  n  ~A  n  ~R  n  T)= 
=  p(I)  +  p(AM)  *  p(~I)  +  p(R/~A  n  ~I)  *  p(~A/~I)  *  p(~I)  + 
+  p(T/~I  n  ~A  n  ~R)  *  p(~R/~A  n  ~I)  *  p(~A/~I)  *  p(~I) 

where: 

p(I)  is  estimated  by  the  number  of  identification  error,  i.e.,  the 
proportion  of  non  identified  objects  or  objects  that  have  been 
erroneously  identified, 

p(~I)  =  1  -  p(I)  is  estimated  by  the  proportion  of  objects  correctly 
identified, 

p(A/~I)  is  estimated  by  the  proportion  of  objects  with  absolute  position 

errors  that  were  not  generated  from  the  identification  errors, 
p(R/~A  n  ~I)  is  estimated  from  the  proportion  of  objects  with  relative 

position  errors  that  were  not  generated  from  the  previous  errors, 
p(~R/~A  n  ~I)  is  estimated  from  the  proportion  of  objects  without 

identification,  absolute  position  and  relative  position  errors, 
p(T/~I  n  ~A  n  ~R)  is  estimate  from  the  proportion  of  objects  with 

topology  errors  that  were  not  generated  from  the  previous  errors. 

CONCLUSIONS 

That  proposed  methodology  for  errors  evalution  in  spatial  database,  aims  at 
reducing  the  time  and  costs  associated  to  qualify  and  validate  that  spatial 
database  for  GIS  users.  Additional  studies  wiil  be  carried  out  to  handle  the 
sporious  polygons  that  can  be  generated  for  each  type  of  errors.  These  polygons 
may  complicate  the  sampling  procedure  and  increase  the  sample  size,  time  and 
costs. 


506 


ACKNOWLEDGMENTS 


The  authors  are  grateful  to  Escola  Politecnica  da  Universidade  de  Sao  Paulo 
which  supported  this  research  and  to  Dr.  Linda  Lee  Ho,  who  supported  the 
statistical  aspects  of  Poisson  multivariate  distribution. 


REFERENCES 

Bartlett,  M  S.  1975.  The  statistical  analysis  of  spatial  pattern.  Chapman  and  Hall, 
London,  90  p. 

Bellhouse,  D.  R.  1988.  Spatial  sampling.  In:  Kotz,  S.;  Johnson,  N.L.  (ed.) 
Encyclopedia  of  Statistical  Science,  New  York,  John  Wiley,  1988.  V.  8,  p.  581- 
4. 

Berger,  T.  1981.  Information  theory  and  coding  theory.  .  In:  Kotz,  S.;  Johnson, 
N.L.  (ed.)  Encyclopedia  of  Statistical  Science.  New  York,  John  Wiley,  V.4,  p. 
125-41. 

Diggle,  P.  J.  1979.  Statistical  methods  for  spatial  point  patterns  in  ecology.  In: 

Cormack,  R.M.;  Ord,  J.K.,  ed.  Spatial  and  temporal  analysis  in  ecology. 

Fairland,  International  Co-operative  Pub.  House,  p. 95-1 50. 
Diggle,  P.  J.  1981.  Some  graphical  methods  in  the  analysis  os  spatial  point 

patterns.  In:  Barnett,  V.,  ed.,  Interpreting  multivariate  data.  Chichester,  John 

Wiley,  chap.  4,  p.  55-73. 
Ho,  L.  L.    1995.  Analysis  of  multivariate  counts.  Sao  Paulo.  109p.  Thesis  (PhD)  - 

Escola  Politecnica,  Universidade  de  Sao  Paulo,  Brazil  (in  Portuguese). 
Jensen,  D.  R.  1985.  Multivariate  distributions.  .  In:  Kotz,  S.;  Johnson,  N.L.  (ed.) 

Encyclopedia  of  Statistical  Science,  New  York,  John  Wiley,  V.6,  p.  43-54. 
Johnson,  N.L.;  Kotz,  S.  1969.  Distributions  in  statistics:  discrete  distributions. 

New  York,  John  Wiley.  Chap.ll:  Multivariate  Discrete  Distributions,  p. 281- 

323. 

Kish,  L.  1965.  Survey  sampling.  Chapter  9:  Area  sampling,  p.  301-358.  New 
York,  John  Wiley. 

Quintanilha,  J.  A.  1996.  Errors  and  uncertainties  in  a  spatial  database  for  GIS 

(non-official  version).  193p.  Thesis  (PhD)  -  Escola  Politecnica,  Universidade  de 

Sao  Paulo,  Brazil  (in  portuguese). 
Ripley,  B.  D.  1981.  Spatial  statistics.  New  York,  John  Wiley  &  Sons,  252  p. 
Ripley,  B.  D.  1988.  Spatial  data  analysis.  .  In:  Kotz,  S.;  Johnson,  N.L.  (ed.) 

Encyclopedia  of  Statistical  Science.  New  York,  John  Wiley,  V.  8,  p.  570-3 
Shaw,  G,  Wheeler,  D.  1985.  Statistical  techniques  in  geography.  S.I.:  Spatial 

indices  &  pattern  analysis,  Chap.  16. 


507 


Assessing  the  Accuracy  of  a 
Regional  Land  Cover  Classification 

William  Clerke1  Raymond  Czaplewski2 
Jeff  Campbell3  and  Janet  Fahringer3 

ABSTRACT 


The  Southern  Region  USDA  Forest  Service  recently  completed  the 
Southern  Appalachian  Assessment  (SAA).  The  Assessment  is  a  broad 
scale  interagency  analysis  and  sharing  of  existing  information  relative  to 
the  natural  and  human  resources  of  the  region.  The  SAA  encompasses 
over  36  million  acres  extending  from  Northern  Virginia  to  Northern 
Alabama.  It  was  clear  early  in  the  Assessment  that  a  uniform  land  cover 
data  layer  would  be  required  to  support  analysis  and  modeling. 

Pacific  Meridian  Resources  (PMR),  in  cooperation  with  the  Forest 
Service,  developed  a  16-class  land  cover  map  of  the  Assessment  area.  The 
classification  included  nine  Anderson  level  3  forest  cover  classes.  Land 
cover  polygons,  based  on  a  minimum  mapping  unit  of  two  acres,  were 
derived  from  the  final  pixel  classification.  This  data  set,  in  raster  and 
vector  formats,  was  provided  to  the  Assessment  participants  as  an  input 
for  deriving  required  products.  The  Forest  Service  took  the  lead  in 
cooperation  with  PMR  in  conducting  an  accuracy  assessment  to  validate 
the  polygon  land  cover  data  theme. 

A  sampling  design  included  aerial  photo  interpretation  and  ground 
examination  phases  was  developed.  During  the  aerial  photo  phase,  data 
was  collected  on  approximately  1500  sample  polygons  within  239 
systematically  distributed  Primary  Sampling  Units  (PSU's).  Sample 
selection  procedures  were  implemented  that  accounted  for  the  skewed 
spatial  and  proportional  distribution  of  the  land  cover  classes  within  the 
Assessment  area. 


1  Remote  Sensing  Specialist,  Southern  Region,  USDA  Forest  Service,  Atlanta,  GA. 

2  Biometricion,  Rocky  Mountain  Forest  and  Range  Experiment  Station,  Fort  Collins,  CO. 

3  Project  Manager  and  Senior  Analyst  respectively,  Pacific  Meridian  Resources,  Portland,  OR. 


508 


Sampling  Satellite  Images  for  Area 
Estimates  in  a  Large  Region 

F.J.  Gallego1 

Abstract.  This  paper  describes  two  alternative  methods  (square  sites  and 
strips)  to  select  a  sample  of  sites  for  satellite  image  analysis  giving 
priority  to  a  particular  type  of  area  (intensive  agricultural  or  forest)  when 
no  suitable  stratification  is  available.  Both  methods  are  based  on  two- 
phase  sampling.  For  square  sites,  a  systematic  sample  is  first  drawn  on 
the  SPOT  reference  grid.  Pre-selected  sites  are  visually  photo-interpreted 
for  a  rough  estimation  of  the  percentage  of  agricultural  land  or  forest, 
that  determines  the  sampling  probability  in  the  second  phase.  The 
alternative  scheme,  based  on  thin  satellite  image  strips,  seems  to  be  more 
efficient  than  sampling  square  sites  according  to  a  simulation  study 
based  on  ground  survey  data  in  Spain. 


INTRODUCTION 

We  consider  here  two  problems  for  which  high  resolution  satellite  images  are 
sampled  because  analysing  a  complete  coverage  of  is  too  expensive: 

•  Area  estimation  of  a  particular  land  cover  (an  annual  crop,  forest,  etc.)  or 
estimation  of  area  change  between  two  dates. 

•  Accuracy  assessment  of  a  land  cover  map  obtained  by  other  means,  for 
example  by  multi-temporal  analysis  of  low  resolution  satellite  images. 

The  Institute  for  Remote  Sensing  Applications  (IRSA)  of  the  European 
Commission  (EC)  is  involved  in  several  projects  where  these  problems  are 
tackled: 

1 .  Rapid  estimates  of  area  change  for  main  crops  in  the  European  Union  (EU)  of 
the  MARS  Project  (Monitoring  Agriculture  with  Remote  Sensing). 

2.  Multi-country  crop  area  estimates  in  central  Europe.  This  is  part  of  the 
MERA-95  Project  (MARS  and  Environmental  related  Applications). 

3.  FIRS  Project  (Forest  Information  from  Remote  Sensing).  Forest  mapping  and 
estimation  in  Europe,  from  the  Atlantic  to  the  Urals  (Kennedy  et  al  1995). 

4.  TREES  Project  (Tropical  Ecosystem  Environmental  observations  by  Satellite): 
Forest  mapping  in  the  inter-tropical  belt  (Malingreau  et  al  1995). 

5.  Estimation  of  crop  area  change  in  countries  of  the  former  Soviet  Union. 


MARS  Project,  IRSA,  JRC,  21020  Ispra  (Va),  Italy 


509 


The  examples  we  present  below  refer  to  items  1  and  2  (MARS  Project), 
where  the  purpose  is  mainly  statistical.  Mapping  is  not  a  priority.  For  FIRS  and 
TREES  (items  3  and  4)  mapping  is  the  main  priority,  but  area  estimation  and 
evaluation  of  mapping  accuracy  are  associated  targets  (Mayaux  and  Lambin, 
1995).  Item  5  refers  to  a  contract  of  the  EC  with  the  French  company  "SCOT 
Conseil",  in  which  IRSA  will  be  responsible  for  site  sampling. 

SAMPLING  SITES  OF  40  KM  x  40  KM 

This  unit  size  is  reasonable  for  multitemporal  analysis  of  high  resolution 
satellite  images.  They  are  large  enough  to  include  a  variety  of  land  cover  types  in 
most  landscapes,  and,  if  conveniently  located,  fit  inside  marketed  images. 

The  "Rapid  Crop  Area  Estimates  in  the  EC"  of  the  MARS  Project. 

Rapid  crop  area  estimates  are 
being  produced  in  the  European 
Union  by  analysis  of  high 
resolution  satellite  images  on  a 
sample  of  sites  of  40  km  x  40  km 
(figure  1).  Estimates  are  computed 
for  inter-annual  variations,  not  for 
a  particular  year  (Pous  et  al,  1995, 
Sharman  and  De  Boissezon,  1992). 
The  method  is  run  on  a  non 
random  sample  of  53  square  sites 
of  40  km  x  40  km  for  EU-12  (The 
12  members  before  1995).  With 
the  3  new  member  states  (Austria,  Figure  1:  Sample  of  53  sites  in  EU-12 

Finland  and  Sweden),  the  number 

of  sites  has  become  60.  Ground  surveys  are  conducted  on  a  sample  of 
approximately  16  segments  of  1400  m  x  1400  m  in  each  site  (Carfagna  1995): 
40  points  per  segment  are  visited  and  a  subsample  of  farms  are  interviewed 
(Gallego  et  al  1994).  Some  problems  arise  to  extrapolate  results  because  of  the 
non-random  character  of  the  sample  of  sites. 

Sampling  Sites  in  Central  Europe 

A  different  approach  has  been  followed  to  draw  a  sample  of  sites  in  Central 
Europe  (Poland,  Czech  Republic,  Slovakia,  Hungary,  Romania  and  Bulgaria)  in 
the  frame  of  the  MERA-95  Project  (MARS  and  Environmental  Related 
Applications).  The  sampling  procedure,  described  in  more  detail  by  Vossen  et  al 
(1995),  has  two-phases: 


510 


First  sampling  phase: 

A  systematic  sample  is  selected  by  blocks  on  the  K-J  indexes  of  the  "SPOT 
reference  grid".  K  is  related  with  the  satellite  orbit,  and  J  refers  to  the  latitude. 
Sites  on  the  border  are  selected  if  more  than  50%  is  inside  the  studied  region.  In 
the  example  of  central  Europe,  58  sites  were  selected  in  this  phase  (figure  2).  The 
sampling  intensity  is  higher  in  the  north  of  the  area  because  the  distance  r\iat(S) 
between  contiguous  satellite  orbits  is  shorter. 

Second  sampling  phase: 

Spot  XS  digital  quick-looks  or  Landsat-TM  photographic  printouts  of  the  pre- 
selected sites  were  analysed  for  a  rough  visual  estimation  As  of  the  percentage  of 
agricultural  land.  40  sites  were  subsampled  without  replacement  with  a  probability 
proportional  to  Ps=  As  x  r\iat(S)  (figure  3).  Estimates  are  computed  with  a  Horvitz- 
Thomson  or  n  estimator  (Sarndal,  1992)  using  1/AS  as  weights  rather  than  1/PS  to 
compensate  the  variable  sampling  rate  in  the  first  phase. 


Figure  2:  Systematic  sample  of  sites  based  on     Figure  3:  Sample  of  sites  for  image  analysis 
the  SPOT  reference  grid. 


SAMPLING  SEGMENTS  IN  SATELLITE  IMAGE  STRIPS 

An  alternative  sampling  scheme  is  currently  evaluated.  This  approach  is 
based  on  long  narrow  sites.  The  site  is  a  strip  as  long  as  the  studied  territory  along 
a  satellite  orbit  track  (figure  4).  The  width  of  the  site  can  be  between  1  km  and 
10  km,  depending  on  the  land  use  category  to  be  estimated.  1  km  or  2  km  can  suit 
for  agricultural  estimates,  but  wider  strips  would  probably  be  more  efficient  in 
forestry.  Different  sets  of  sites  can  be  defined  for  different  satellite  sensors,  such 
as  Landsat-TM,  SPOT-XS,  or  IRS-1C. 


511 


Sampling  Segments  inside  the  Strips 


Segment  sampling  inside  the  strips  follows  a  two-phase  procedure,  similar  to 
the  site  sampling  scheme  of  the  previous  section.  The  first  phase  can  be 
systematic  or  random;  in  the  simulations  presented  below  we  have  used 
systematic  sampling,  that  seems  to  be  generally  more  efficient  in  a  geographic 
context  (Dunn  and  Harrison,  1993).  In  the  second  phase,  a  subsample  is  selected 
with  a  probability  proportional  to  an  index  of  agricultural  intensity  (or  forest 
intensity),  and  to  the  distance  between  satellite  orbits. 

Image  Acquisition 

One  important  condition  for  the  strip  approach  to  be  efficient  is  that  images 
are  marketed  by  continuous  strips  instead  of  approximately  square  scenes.  In  the 
case  of  Landsat  TM,  the  delivered  image  for  agricultural  purpose  would  have  for 
example  a  size  of  2000  km  x  3  km  =  40  Mbytes  instead  of  the  standard  frames 
180  km  x  180  km  =  200  Mbytes. 

Strips  may  be  divided  into  pieces  corresponding  to  different  receiving 
stations:  During  the  agricultural  season,  image  strips  would  be  delivered 
regardless  of  the  cloud  coverage.  This  gives  in  Europe  approximately  1 1  TM 
images  or  6-7  SPOT  images  per  year.  The  selection  of  images  that  can  be 
exploited  for  land  use  identification  in  a  particular  segment  would  be  made  later 
by  the  photo-interpreter. 


The  target  is  getting  an 
"exploitable  sample"  of  at  least  n 
segments.  For  annual  crops,  a  possible 
definition  of  "exploitable  segment" 
can  be:  a  segment  for  which  at  least  3 
cloud  free  images  are  available 
during  two  successive  agricultural 
seasons  with  a  time  interval  of  at  least 
one  month.  The  definition  of 
exploitable  segment  is  to  be  adapted 
to  each  case;  for  many  forest 
applications,  one  cloud  free  image 
may  be  enough.  The  required  number 
of  second  phase  sample  segments  to 
reach  the  targeted  number  of 
exploitable  segments  is  being  studied. 


Figure  4:  Sites  on  Landsat  tracks 


512 


COMPARING  THE  EFFICIENCY  OF  SAMPLING  STRATEGIES. 


We  do  not  have  yet  data  to  test  the  efficiency  of  the  strip  approach  with  real 
strips  following  satellite  tracks,  but  we  can  simulate  the  scheme  with  data  from 
ground  survey  segments.  The  set  of  data  used  for  this  test  comes  from  1992  and 
1993  ground  surveys  on  8023  square  segments  of  49  ha  from  the  Spanish  Ministry 
of  Agriculture  in  an  area  of  270.000  Km2  (Castilla  y  Leon,  Madrid,  Castilla-la 
Mancha  and  Andalucia).  The  set  of  8023  segments  was  obtained  by  systematic 
sampling  with  3  replicates  in  blocks  of  10  km  x  10  km.  The  sample  is  described 
by  MAPA  (1990),  Ambrosio  (1993),  Fuentes  (1994)  and  Gallego  (1995).  We 
consider  this  set  of  8023  segments  as  population  from  which  we  select  samples  of 
80  segments  with  different  strategies  (figure  4).  Unequal  probability  subsampling, 
when  applied,  was  made  using  1992  data  to  compute  agricultural  intensity 
indexes.  We  compare  area  estimates  in  1993  or  area  change  estimates  1993-1992. 
For  each  strategy,  we  have  repeated  the  procedure  100  times  in  order  to  estimate 
variances  without  unstability  problems,  that  are  well  known  for  the  variance 
computation  of  the  Horvitz-Thompson  estimator  (Cochran,  1977,  Sarndal,  1992). 

Sampling  Strategies  Compared 

•  Random:  simple  random  sampling  without  replacement.  This  is  the  reference 
to  compute  the  relative  efficiency  of  the  other  strategies  as  ratio  of  variances. 

•  Systematic:  Systematic  sampling  ordering  population  segments  by  geographic 
co-ordinates  (latitude  and  then  longitude).  Uniform  probability. 

•  Proportional:  Sample  with  a  probability  proportional  to  arable  land  in  1992. 

•  Sites  systematic:  Systematic  sample  by  clusters  of  16  segments  approximately 
arranged  in  squares  of  40  km  x  40  km  .  Uniform  probability. 

•  Sites  probabilistic  :  The  same  as  the  previous  one  with  probability  proportional 
to  the  arable  land  in  1992.  This  approach  nearly  coincides  with  the  one 
described  above,  that  has  been  proposed  for  the  MERA-95  Project. 

•  Strips:  Two-stage  sampling:  systematic  sampling  of  strips  100  km  apart  from 
each  other  and  sub-sample  with  probability  proportional  to  1992  arable  land. 


Proportional  Sites  probabilistic  Strips 

Figure  4:  Samples  with  different  strategies 


513 


Comments  on  Simulation  Results. 

Some  provisional  conclusions  can  be  drawn  by  examining  Table  1,  that  gives 
the  relative  efficiency  of  the  studied  strategies  compared  with  simple  random 
sampling  for  the  main  annual  crops  and  fallow  (understood  as  arable  land  under 
repose  in  the  crop  rotation). 

•  Systematic  sampling  looks  slightly  superior  to  random  sampling,  but  there  are 
some  important  exceptions. 

•  Sampling  unclustered  segments  with  probability  proportional  to  an  index  of 
agricultural  intensity  (strategy  Proportional)  is  much  more  efficient  (ratio 
ranging  between  2  and  9)  than  simple  random  or  systematic  sampling. 
However  this  strategy  may  be  unsuitable  for  satellite  image  acquisition. 

•  Sampling  by  square  sites  of  40  km  x  40  km  reduces  efficiency  both  with 
systematic  and  unequal  probability  sampling  (Sites  systematic  compared  with 
Systematic  or  Sites  proportional  compared  with  Proportional ).  In  the  case  of 
systematic  sampling,  efficiency  ratios  become  close  to  1  when  area  changes 
are  estimated  instead  of  crop  area  for  a  particular  year.  For  the  more  efficient 
unequal  probability  sampling,  sites  of  40  km  x  40  km  still  have  a  poor 
behaviour. 

•  Sampling  by  strips  is  not  as  good  as  sampling  unclustered  segments,  but  is 
more  efficient  than  sampling  by  square  sites  and  is  compatible  with  satellite 
image  acquisition. 

•  The  efficiency  gain  with  a  two-phase  sampling,  where  the  second  phase  is  a 
subsampling  with  a  probability  proportional  to  arable  land  is  higher  than  the 
efficiency  obtained  in  most  European  countries  with  area  frame  sampling 
stratification  (Taylor  et  al  1996). 

•  Some  exceptions  to  the  comments  above  can  be  noted  for  crops  with  a 
relatively  small  area,  as  we  can  check  in  the  case  of  sugar  beet. 


Table  1:  relative  efficiency  compared  with  random  sampling 


Systematic 
uniform 

Proportional 

Sites 
systematic 

Sites 
probabilistic 

Strips 

Wheat  93 

1.20 

6.53 

0.36 

1.14 

2.80 

Barley  93 

1.61 

6.65 

0.46 

1.12 

4.93 

Cereals  93 

1.42 

7.07 

0.43 

1.47 

5.80 

Sunflower  93 

1.26 

3.53 

0.20 

0.81 

1.69 

Sugar  beet  93 

1.14 

2.85 

0.57 

0.66 

1.69 

Fallow  93 

1.05 

3.96 

0.49 

0.49 

2.11 

A  Wheat 

0.82 

2.92 

1.36 

1.20 

1.95 

A  Barley 

0.80 

2.43 

1.01 

0.80 

2.13 

A  Cereals 

1.11 

2.39 

1.23 

0.85 

1.94 

A  Sunflower 

1.13 

2.47 

1.13 

1.54 

2.09 

A  Sugar  beet 

0.87 

3.06 

2.42 

2.74 

1.90 

A  Fallow 

1.00 

3.09 

0.91 

0.75 

2.27 

514 


CONCLUSIONS 


Two-phase  sampling  may  provide  a  scheme  for  an  adequate  use  of  remote 
sensing  applied  to  land  cover  area  estimation  in  large  regions,  specially  if  no 
suitable  stratification  is  available  or  if  available  stratification  is  not  very  efficient. 

Image  acquisition  by  long  thin  strips  instead  of  approximately  square  scenes 
can  improve  the  efficiency  of  associated  sampling  plans.  However  many  factors 
are  still  to  be  studied,  including  image  distribution  policy,  cost  of  ground  surveys 
and  image  analysis,  and  efficiency  ratio  for  different  land  cover  types  in  different 
landscapes. 

ACKNOWLEDGEMENTS 

We  are  grateful  to  the  Spanish  Ministry  of  Agriculture,  in  particular  to  Porfirio 
Sanchez,  Jose  Maria  Fernandez  del  Pozo,  and  Maria  Jose  Postigo,  that  has  kindly 
provided  data  to  test  the  efficiency  of  sampling  strategies. 

REFERENCES 

Ambrosio  L.,  Alonso  R.,  Villa  A.,  1993,  Estimation  de  superficies  cultivadas  por 
muestreo  de  areas  y  teledeteccion.  Precision  relativa.  Estadistica  Espanola, 
Vol.  35,  pp.  91-103. 

Carfagna  E.,  Gallego  F.J.,  1995,  Yield  Estimates  from  Area  Frame  at  European 
Level.  Seminar  on  Yield  Forecasting.  Villefranche  sur  Mer,  24-27  October 
1994.  pp.  237-240. 

Cochran  W.,  1977,  Sampling  Techniques.  New  York:  John  Wiley  and  Sons 
Dunn  R.,  Harrison  A.R.,  1993,  Two-dimensional  systematic  sampling  of  land  use. 

JRSS:  Applied  Statistics,  vol.  42  n.  4,  pp.  585-601. 
Fuentes  M.,  Gallego  F.J.,  1994,  Cluster  area  estimation  on  a  sample  by  square 

grid,  getting  a  stable  variance  for  the  survey  of  the  Spanish  Ministry  of 

Agriculture.  Conference  on  the  MARS  Project:  overview  and  perspectives. 

Office  for  Publications  of  the  E.C.  Luxembourg,  pp.  146-149. 
Gallego  F.J.,  Delince  J., Carfagna  E.,  1994,  Two-Stage  Area  Frame  Sampling  on 

Square  Segments  for  Farm  Surveys.  Survey  Methodology  ,  vol  20,  No.  2  ,  pp. 

107-115. 

Gallego  F.J.,  1995,  Sampling  Frames  of  Square  Segments,  Report  EUR  16317, 
Office  for  Publ.  of  the  E.C.  Luxembourg.  68  pp.  ISBN  92-827-5106-6 


515 


Kennedy  P.,  Folving  S.,  Mc.  Cormick  N.,  1995,  European  Forest  Ecosystems 
Mapping  and  Forest  Statistics:  the  FIRS  Project.  Proceedings  of  the  wokshop 
on  "Defining  a  System  of  Nomenclature  for  European  Forest  Mapping.  June 
13-15,  1994,  Joensuu,  Finland,  pp.  33-48.  Office  for  Pubblications  of  the  EC. 
Luxembourg. 

Malingreau  J.P.,  Achard  F.,  D'Souza  G.,  Stibig  H.T.,  D'Souza  J.,  Estreguil  C, 
Eva  H.,  1995,  AVHRR  for  Global  Tropical  Forest  Monitoring:  The  Lessons  of 
the  TREES  Project.  Remote  Sensing  Reviews,  1995,  Vol.  12,  pp.  29-40. 

MAPA,  1990,  El  Marco  de  Areas  como  Instrumento  de  base  para  la  Estadfstica  de 
superficies  de  cultivo.  Boletm  Mensual  de  Estad.  Min.  Agric.  n.  12,  pp.  85- 
106.  Madrid 

Mayaux  Ph.,  Lambin  E.F.,  1995,  Estimation  of  tropical  forest  area  from  coarse 
spatial  resolution  data:  a  two-step  correction  function  for  proportional  errors 
due  to  spatial  aggregation.  Remote  sens.  Environ,  n.  53,  pp  1-15 

Pous  B.,  Reboux  D.,  Guerif  M.,  Delecolle  R.,  Fischer  A.,  de  Boissezon  H.,  1995. 
Previsions  du  rendement  et  information  satellitaire:  acquis  de  Taction  IV  du 
projet  MARS  et  nouvelles  perspectives.  Seminar  on  Yield  Forecasting. 
Villefranche  sur  Mer,  24-27  October  1994.  pp.  567-578. 

Sarndal  C.E.,  Swenson  B.,  Wretman  ,  1992,  Model  Assisted  Survey  Sampling. 
Springer  Verlag. 

Sharman  M.,  de  Boissezon  H.,  (1992),  Action  IV,  de  l'image  aux  statistiques: 
bilan  operationnel  apres  deux  annees  d'  estimation  rapides  des  superficies  et 
des  rendements  potentiels  au  niveau  Europeen.  Conference  on  the  Appl.  of 
Remote  Sensing  to  Agricultural  Statistics  (Belgirate).  Office  for  Publications  of 
the  E.C.  Luxembourg,  pp.  177-186. 

Taylor  J.,  Sannier  C,  Delince  J,  Gallego  F.J.,  1996,  Regional  crop  inventories  in 
Europe  assisted  by  remote  sensing  (in  press). 

BIOGRAPHICAL  SKETCH 

Javier  Gallego  graduated  in  Mathematical  Statistics  in  Valladolid  (Spain)  and 
made  a  PhD.  in  Multivariate  Descriptive  Analysis  in  Paris.  He  taught  Applied 
Statistics  again  in  Valladolid,  where  he  became  Professor  in  1986.  Then  he 
decided  to  flee  the  University  and  join  the  European  Commission.  Since  1988  he 
works  in  the  Joint  Research  Centre  in  Ispra  (Italy),  where  he  is  in  charge  of 
statistical  aspects  of  the  MARS  Project  (Monitoring  Agriculture  with  Remote 
Sensing). 


516 


Cooperative  Accuracy  Assessment 
Strategies  for  Sampling  a  Natural 
Landcover  Map  of  Arkansas 

Robert  S.  Dzur1,  Michael  E.  Garner,  Kimberly  G.  Smith,  W. 
Fredrick  Limp,  Donald  G.  Catanzaro,  and  Richard  L.  Thompson 


Abstract.-The  National  Biological  Service's  Gap  Analysis  Program 
mandates  mapping  actual  vegetation  on  a  state-wide  basis.  Currently, 
there  are  33  ongoing  or  completed  state  gap  analysis  programs.  As  those 
vegetation  layers  are  produced  for  each  state,  an  increasing  concern  and 
need  exists  to  supply  some  measure  of  map  accuracy.  However, 
challenges  associated  with  designing  and  implementing  systematic 
validation  of  state-wide  natural  landcover  maps  are  significant  and 
potentially  cost  prohibitive. 

A  "Rapid  Assessment  Track"  (RAT)  for  systematic  accuracy  assessment 
of  a  Landsat  Thematic  Mapper  (TM)  based  natural  landcover  map  of 
Arkansas  was  developed  and  implemented.  Arkansas  gap  analysis 
adopted  a  stratified  random  sampling  strategy  designed  to  minimize  field 
data  collection  costs.  Comprehensive  mapping  goals  of  the  Arkansas  gap 
analysis  project  warranted  development  and  strong  reliance  upon  state 
partnership  networks.  Arkansas  Forestry  Commission  (AFC)  personnel 
conducted  actual  ground  data  collection  of  the  systematic  sampling 
network.  AFC  provided  an  established  and  capable  field  apparatus  with 
state-wide  management  authority  to  collect  required  validation  data  in  a 
timely  manner  at  reduced  cost.  Matching  the  comprehensive  Arkansas 
gap  analysis  mapping  objectives  with  an  equally  comprehensive  state- 
wide accuracy  assessment  would  not  have  been  feasible  without  the 
cooperative  partnership  between  AFC  and  Arkansas  gap  analysis. 


INTRODUCTION 

Gap  analysis  combines  state-wide  Landsat  TM  derived  natural  landcover  data 
with  predicted  terrestrial  vertebrate  distributions  to  model  biological  diversity  at 
eco-regional  scales  (Scott  et  al.  1993).  Under  this  scenario,  it  is  important  to 
evaluate  accuracy  of  thematic  information,  upon  which  models  and  potential 
management  decisions  may  be  based. 

Error  matrices  are  one  commonly  used  method  for  assessment  of  landcover 


1 .  Address  correspondence  to:  CAST,  University  of  Arkansas,  Ozark  Hall  Room  1 2, 
Fayetteville,  AR  72701  (rob@cast.uark.edu) 


517 


maps  produced  by  analysis  of  remotely-sensed  digital  multispectral  satellite  data 
(Congalton  1988,  Lark  1995,  Verbyla  and  Hammond  1995).  The  error  matrix  is  a 
two-way  contingency  table  that  compares  classified  pixels  to  a  set  of 
independently  collected  reference  data  (Janssen  and  van  der  Wei  1994). 

Collecting  unbiased  independent  reference  samples  across  a  13,798,200  ha 
study  area  such  as  Arkansas  becomes  problematic.  Because  of  their  size,  state- 
wide natural  landcover  maps  offer  many  difficult  logistical  issues  for  systematic 
thematic  accuracy  assessments.  Among  these  issues  are  limited  access  to  private 
lands,  time,  and  cost  constraints.  These  broad  issues  characterize  the  problem  of 
implementing  an  accuracy  assessment  on  a  state-wide  scale.  While  those  issues 
are  readily  apparent  and  relatively  easy  to  isolate,  the  challenge  remains  to  devise 
available  options  that  address  those  broad  constraints.  The  Arkansas  gap  analysis 
project  required  an  implementation  plan  -  a  "Rapid  Assessment  Track"  (RAT)  -  for 
gap  analysis  landcover  map  accuracy  assessment  that  minimized  direct  data 
collection  time,  costs,  and  land  access  limitations. 

Perhaps  the  most  influential  factor  on  the  magnitude  of  logistical  obstacles  lies 
in  the  character  of  the  sample  design  used  for  accuracy  assessment.  Janssen  and 
van  der  Wei  (1994)  suggested  that  poor  accessibility  and  limited  budgets  may 
result  in  a  cluster-based  sampling  approach.  While  clustered  sampling  might 
reduce  the  cost  to  travel  between  sample  units,  field  crews  still  needed  to  visit  and 
record  sample  site  information.  Field  crews  for  accuracy  assessment  were 
solicited  from  Arkansas  gap  analysis  cooperators  gathered  at  the  third  annual  state- 
wide gap  analysis  meeting  in  February  of  1995  in  Little  Rock  (Dzur  et  al.  In 
press).  Here,  Arkansas  Forestry  Commission  (AFC)  demonstrated  a  willingness  to 
participate  in  the  RAT  initiative. 

As  the  state  agency  responsible  for  the  protection  and  development  of 
Arkansas'  forest  resources,  AFC  possessed  key  qualities  needed  to  implement 
RAT.  AFC  provided  a  well-developed  organizational  structure.  Furthermore, 
AFC  provided  a  knowledgable  and  well-equipped  (vehicles,  etc.)  personnel  base 
trained  in  land  navigation  and  natural  landcover  recognition.  This  state-wide 
personnel  network  facilitated  simultaneous  sample  data  collection  across  the  entire 
state,  thus  minimizing  collection  time. 

This  paper  presents  development  of  cooperative  implementation  strategies  for 
systematic  accuracy  assessment  of  a  natural  landcover  map  of  Arkansas.  Research 
objectives  described  include:  automated  sample  design  development  and 
automated  field  material  production  directed  toward  realization  of  a  simple, 
systematic,  and  cost  effective  state-wide  accuracy  assessment  approach  that  relied 
upon  an  established  partnership  between  the  Center  for  Advanced  Spatial 
Technologies  (CAST)  and  AFC. 

METHODS 

RAT  operational  viability  resulted  from  a  carefully  devised  methodology  related 
to  sample  design  and  field  material  composition. 


518 


Sample  Site  Selection 


A  critical  RAT  objective  was  to  employ  random  sampling  at  clustered  locations. 
This  objective  framed  another  compelling  reason  to  involve  AFC  in  the  sampling 
effort.  Commanding  state-wide  presence  and  authority,  AFC  field  crews  gained 
ready  access  to  lands  otherwise  inaccessible  due  to  the  geography  of  land 
ownership.  AFC  participation  helped  enable  a  random  site  selection  design. 

Site  Selection  Processing 

Sample  point  selection  methods  were  adopted  and  modified  from  Utah  gap 
analysis  (Edwards  et  al.  1995).  Arkansas  gap  analysis  spectral  classification 
procedures  utilized  National  Resource  Conservation  Service  (NRCS)  State  Soil 
Geographic  Data  Base  (STATSGO)  for  stratification  of  spectral  space. 
Accordingly,  RAT  sampling  was  stratified  by  major  landforms  of  Arkansas  (figure 
1).  U.  S.  Geological  Survey  (USGS)  7.5  minute  quadrangles  served  as  the  first 
filtering  strata  for  the  clustering  of  random  samples.  A  total  of  100  USGS  7.5 
minute  quadrangles  (figure  2)  were  randomly  selected  in  proportion  to  the  relative 


Figure  1.  Major  Arkansas  landforms.  Figure  2.  Sample  quadrangle  locations. 

area  of  Arkansas'  10  major  landforms.  For  example,  since  the  Arkansas  River 
Valley  and  Ridge  landform  region  comprised  10%  of  the  state,  this  region  was 
randomly  assigned  10  quadrangles. 

Once  identified,  quadrangles  were  further  subdivided  into  two  road-based  strata 
within  which  a  total  of  20  1-ha  sample  points  per  quadrangle  were  distributed. 
Each  quadrangle  was  buffered  by  360  m  to  ensure  that  roads  on  surrounding 
quadrangles  did  not  influence  random  site  selection.  Then,  the  1:100,000  scale 
USGS  Digital  Line  Graph  (DLG)  road  network  was  buffered  by  300  m.  Within 
this  buffer  zone,  10  sites  were  randomly  selected.  Next,  10  additional  sites  were 
systematically  located  adjacent  to  one  another  based  upon  a  single  randomly 
placed  starting  point  selected  outside  the  roads  buffer.  These  10  off-road  sites 
were  positioned  so  as  not  to  touch  the  roads  buffer  and  to  lie  on  a  1  km  transect 
oriented  north  for  easy  compass  location  by  the  ground  crew.  A  GRASS  GIS 
UNIX  script  automated  random  site  selection  processing  for  the  100  quadrangles. 

Field  Material 

While  AFC  provided  the  labor  force  required  to  visit  those  sample  sites,  CAST 
needed  to  ensure  a  quality  data  collection  effort.  Accurately  locating  sample 


519 


points  in  the  field  was  the  central  concern  and  most  difficult  expectation.  Since  it 
was  impractical  to  equip  AFC  with  Global  Positioning  System  (GPS)  equipment, 
an  alternative  was  to  provide  AFC  crews  with  clearly  delineated  and  detailed  field 
map  sheets.  Designing  the  most  effective  field  map  for  AFC  personnel  required 
intimate  understanding  of  AFC  land  navigation  training  guidelines.  AFC  outlined 
these  guidelines  and  recommended  a  ranked  priority  of  routine  AFC  land  reference 
features  for  inclusion  on  field  maps.  Since  the  chain  (20  m)  is  the  fundamental  unit 
of  field  measurement  used  by  AFC  personnel,  AFC  requested  a  chain  map  scale. 
In  keeping  with  AFC  needs,  maps  were  printed  with  a  mile  barscale  subdivided  by 
tenths  (80  chains  =  1  mile).  According  to  AFC,  Public  Land  Survey  System 
(PLSS)  line  work  was  an  absolute  requirement.  Once  PLSS  data  became  available 
from  USGS,  CAST  manufactured  a  field  data  collection  package  complete  with 
detailed  field  maps,  map  classification  keys,  and  disposable  cameras. 

Field  Map  Production 

Two  sets  of  8.5  x  11  field  maps  (one  color  and  one  black  and  white)  were 
automatically  generated  for  each  quadrangle  using  the  Interactive  Mapper 
(www.cast.uark.edu/products/MAPPER/)  GRASS  script  modified  to  print  field 
map  specific  features.  Distinct  color,  weight,  and  style  symbology  were  used  to 
maximize  contrast  between  map  features.  Separate  background  raster  maps  were 
rendered  for  color  and  black  and  white  maps.  A  Landsat  TM  unsupervised 
classification  with  tasseled-cap  (Crist  and  Cicone  1984)  pseudo  color  table  (RGB 
equals  wetness,  greenness,  brightness)  was  used  as  the  background  raster  for  color 
maps.  Figure  3,  for  example,  depicts  the  landscapes  of  the  St.  Francis  National 
Forest  (upper  right)  on  Crowley's  Ridge  and  the  Mississippi  alluvial  valley. 
Topologically  Integrated  Geographic  Encoding  and  Referencing  System  (TIGER) 
water  was  used  as  the  background  raster  layer  for  black  and  white  maps  (figure  4). 

Numerous  1:100,000  scale  reference  layers  were  then  printed  over  the  raster 
background.  These  reference  layers  appeared  on  both  map  sets  and  included: 
TIGER  roads  (with  varying  line  weights  and  styles  for  distinct  road  classes), 
TIGER  railroads,  DLG  hydrology,  DLG  contours,  and  DLG  PLSS  data  (with 
labelled  sections).  USGS  Geographic  Names  Information  System  (GNIS) 
populated  place  names  and  their  site  locations  were  printed  and  oriented  such  that 
the  names  would  always  be  contained  entirely  within  the  map  area.  A  white  dot 
surrounded  by  a  black  ring  marked  the  20  sites  obtained  from  the  site  selection 
procedure.  Sites  were  given  an  alphabetical  (A-K)  letter  designation  from  north  to 
south  where  K  always  indicated  the  1  km  transect.  Margin  information  contained 
barscale,  state  locator  map,  north  arrow,  quadrangle  name  and  unique  number 
identifier,  regional  coverage  (Latitude/Longitude),  source  data,  and  CAST  contact 
information. 

A  total  of  300  (200  color,  100  black  and  white)  maps  were  delivered  to  AFC. 
One  set  of  color  maps  was  maintained  at  AFC  headquarters  in  Little  Rock,  AR. 
The  remaining  maps  were  bundled  by  AFC  district  for  distribution  to  district 
foresters  and  then  to  actual  field  crews.  To  avoid  potential  bias  in  data  collection, 


520 


Figure  3.  Color  field  map. 


521 


Figure  4.  Black  and  white  field  map. 

no  explanation  or  meaning  was  attributed  to  the  pattern  of  color  portrayed  by  the 
Landsat  TM  background  layer.  The  black  and  white  map  doubled  as  field  data 
sheet.  On  the  reverse  of  this  map  were  20  data  entry  rows  beginning  with  sample 
site  A.  Data  entry  columns  included:  date,  landcover  code,  photo  #,  source,  and 
comments.  While  on-site  visits  were  preferred,  it  was  recognized  that  some  sites 


522 


may  have  been  previously  inventoried.  Therefore,  the  source  column  listed  three 
data  survey  alternatives:  on-site  visit  (V),  inventory  records  (I),  or  personal 
communication  (P).  At  the  top  of  the  field  data  sheet,  name  and  phone  number  of 
the  field  observer  were  recorded.  Upon  RAT  completion,  black  and  white  map/ 
field  data  sheets  were  returned  to  CAST. 

Classification  Key  Development 

A  UNESCO/TNC  based  vegetation  classification  scheme  was  adopted  for 
Arkansas  gap  analysis  landcover  mapping  (Foti  et  al.  1994).  The  UNESCO/TNC 
classification  system  is  characterized  by  its  hierarchical  structure  (Jennings  1993). 
To  help  field  crews  match  ground  data  to  this  classification  system  a  coded  semi- 
dichotomous  key  was  assembled  from  the  classification  system  (figure  5).  The  key 
was  designed  to  lead  the  user  through  a  possible  35  main  entries  with  multiple 
contrasting  sub-entries  defining  particular  landcover  criteria  to  identify  the  correct 
landcover  map  class.  Flowing  linearly  in  accordance  with  the  hierarchical 
structure  of  the  UNESCO  based  classification,  the  key  first  delineates  class 
membership  at  a  system  level  (e.g.,  Terrestrial,  Developed  Cover,  Palustrine, 
Riverine,  etc.).  The  two-page  (double-sided)  key  reads  from  left  to  right  starting 


CODE 


DEFINITION 


GAP  CLASS 


GO  TO 


1(a) 
(b) 
1(c) 
2(a) 
2(b) 
3(a) 
3(b) 


4(a) 
4(b) 

4(c) 

4(d) 

4(e) 

5(a) 
5(b) 

6(a) 
6(b) 

7(a) 
7(b) 
7(c) 


Primarily  water  (bottomlands) 

Primarily  land 

Primarily  altered  by  humans 

Long  term  to  permanent  flooding. 

Seasonal  flooding  (short  to  long  duration). 

Streamside  areas  and  channels. 

Wetlands  and  deep  water  habitat.  30%  tree  coverage 

lacking  trees  and  shrubs.  Total  area  usually  >  20  acres. 

TERRESTRIAL 


TERRESTRIAL 
URBAN/ AGRI/REGEN 

PALUSTRINE 
RIVERINE 

LACUSTRINE 


Areas  dominated  by  trees  with  a  total  canopy  cover 
of  61%  or  more,  tree  crowns  usually  interlocking. 
Areas  dominated  by  trees  with  a  total  canopy  cover  of  26-60%, 
most  tree  crowns  not  touching  each  other.  An  herbaceous 
or  shrub  understory,  or  both  are  usually  present. 
Areas  dominated  by  shrubs  that  generally  exceed  0.5 
meters  in  height  when  mature  with  a  total  canopy  cover 
of  26%  or  more.  A  tree  canopy  cover  of  26%  or  less 
may  be  present. 

Areas  dominated  by  grass,  grass-like,  or  forb  vegetation 
with  a  tree  or  shrub  component  not  exceeding  26%  cover. 


FOREST 


WOODLAND 


SHRUBLAND 


Areas  where  vegetation  cover  is  less  than  5%. 

Mainly  evergreen  forest  (>  than  75%  evergreen  trees) 
Mainly  deciduous  or  mixed  (25%  -  75%  evergreen  trees) 

Dominated  by  eastern  redcedar  {Juniperus  virginiana) 
Dominated  by  pines 

Shortlcaf  pine  (Pinus  echinata)  dominant 

Loblolly  pine  (Pinus  taeda)  dominant 

Mixed  shortlcaf  pine  -  loblolly  pine  (sometimes  hickory) 


GRASSLAND 
BARE 


T.l.a.9.c.I 


T.l.A.9.b.I 

T.l.A.9.b.II 

T.l.A.9.b.III 


2 
4 

35 
3 

23 
32 

34 


5 

15 

17 

18 

22 

6 
8 


Figure  5.  Example  page  from  Arkansas  gap  analysis  classification  key. 


523 


with  landcover  code,  followed  by  a  code  description,  the  equivalent  gap  analysis 
classification  notation  (e.g.,  T.1.A.9.D.I  -  Pinus  echinata),  and  finally  a  "Go  To" 
column  with  a  number  designating  the  next  step  down  the  hierarchy  toward  a  level 
five  (cover  type)  classification. 

To  locate  shortleaf  pine  (Pinus  echinata)  dominated  forest  with  the  key,  for 
example,  AFC  field  personnel  start  with  code  la.  As  la  -  "primarily  water 
(bottomlands)"  does  not  apply,  proceeding  to  lb  -  "primarily  land"  satisfies  the 
present  conditions  and  points  to  4.  The  first  entry  (4a),  "areas  dominated  by  trees 
with  a  total  canopy  cover  of  61%  or  more,  tree  crowns  usually  interlocking",  fits 
the  definition  of  forest  leading  the  search  to  5.  Again  the  first  entry  (5a),  "mainly 
evergreen  forest  (>  than  75%  evergreen  trees)",  corresponds  to  ground  evidence 
and  leads  to  6.  As  6a,  "dominated  by  eastern  redcedar  (Juniperus  virginiana)", 
does  not  agree,  the  next  alternative,  (6b),  "dominated  by  pines",  leads  to  7  where 
7a,  "Shortleaf  pine  (Pinus  echinata)  dominant",  matches  the  site  characteristics. 
Then,  7a  was  recorded  on  the  field  data  sheet.  Instructions  to  field  crews  were  to 
record  the  code  about  which  they  had  the  most  confidence. 

Site  Verification  and  Pictorial  Key  Concept 

When  landcover  codes  were  listed  on  the  field  data  sheet,  a  fundamental 

question  remained.  What  did  Pinus  echinata  (7a)  look  like?  Since  data  were 

collected  by  AFC  field  crews  rather  than  CAST  staff,  site  characteristics  were 

essentially  unknown.  To  supplement  the  recorded  codes,  25  of  the  100  sample 

quadrangles  were  arbitrarily  allocated  a  disposable  Kodak  Fun  Camera.  At  least  2 

quadrangles  in  each  major  landform  unit  were  assigned  cameras.  For  each  sample 

location,  field  crews  took  one  picture  (figure  6)  that  best  captured  the  landcover 

code  and  recorded  the  corresponding 

exposure  number  in  the  photo  #  column. 

To  track  camera  and  quadrangle,  a 

CAST  address  label  printed  with 

quadrangle  name  and  number  was 

attached  to  each  camera.  Using  RAT  as 

a  foundation,  these  photos  might  also 

provide  the  initial  basis  for  a  Web 

enabled  multiscale  pictorial  key  of  the 

Arkansas  gap  analysis  classification  that 

ij  .  i       •    .  Figure  6.  Photo  from  site  H  (figure  4). 

would  support  written  descriptions  6  6 

with  pictorial  examples  of  landcover  map  categories  using  ground  based  photos, 

aerial  videography  frames,  and  satellite  imagery. 

RESULTS 

A  total  of  90  field  data  sheets  (quadrangles)  were  returned.  From  those 
quadrangles,  AFC  collected  a  total  of  1777  field  sample  points.  Two  quadrangles 
lacked  the  off-road  points  because  they  were  not  printed  on  the  field  maps.  Only  3 


524 


sites  were  not  collected  because  the  landowner  would  not  allow  the  AFC  crew  to 
enter  the  property.  Of  the  total  sample  sites,  1601  were  actual  on-site  visits,  156 
were  based  on  inventory  records,  and  20  were  based  on  personal  communication. 

A  total  of  23  cameras  were  returned.  From  those  cameras,  412  photos  were 
developed.  Although  every  roll  was  not  completely  developed,  most  photos 
appeared  very  informative.  Even  underexposed  photos  provided  informational 
clues  related  to  canopy  closure.  Expected  benefit  relative  to  a  total  cost  of  $330  for 
camera  purchase  and  film  developing  justified  the  experimental  camera  concept. 

CONCLUSIONS 

RAT's  primary  accomplishment  was  implementing  a  cooperative  state-wide 
systematic  data  collection  effort.  As  noted  in  the  methods  section,  two  major 
elements  guided  successful  RAT  implementation.  First,  the  sample  design 
achieved  immediate  reductions  in  data  collection  time  through  stratification  on 
quadrangles  and  roads.  Second,  clear  and  concise  field  materials  supported  AFC 
field  observations.  Field  materials  were  generated  to  assist  sample  site  location, 
description,  and  documentation.  Furthermore,  field  materials  formed  the  vital  link 
between  sample  design  and  field  data  collection. 

As  potential  users  of  the  Arkansas  natural  landcover  map,  AFC  recognized 
legitimate  value  in  RAT  objectives  and  contributed  a  proportional  commitment  to 
the  successful  completion  of  RAT.  Although  AFC's  participation  in  RAT 
minimized  direct  costs  to  the  Arkansas  gap  analysis  project,  AFC  costs  were  not 
minimal.  In  fact,  Robert  McFarland,  AFC  Assistant  State  Forester,  estimated 
manpower  costs  alone  at  one  half  man  year  (over  $10,000).  Still,  RAT 
demonstrated  that  partnerships  strengthened  by  well-developed  planning  can 
overcome  major  impediments  to  state-wide  accuracy  assessment  of  land  access, 
collection  time,  and  cost. 

ACKNOWLEDGMENTS 

The  described  assessment  technique  was  funded  through  grants  from  the  United 
States  Fish  and  Wildlife  Service,  Award  #14-16-0009-1567  RWO  #13,  as  a  part  of 
the  National  Gap  Analysis  Program,  and  the  National  Aeronautics  and  Space 
Administration,  Award  #  NAGW-41 17  with  voluntary  contributions  from  AFC. 
CAST  is  appreciative  of  all  AFC  field  crews  that  participated  in  this  project. 
Thanks  is  due  to  Robert  McFarland  (AFC  Assistant  State  Forester)  and  J.  Garner 
Barnum  (AFC  Assistant  State  Forester)  for  coordinating  RAT  at  AFC  and 
contributing  their  thorough  insight  of  AFC  operations  to  RAT  planning. 

REFERENCES 

Congalton,  C.G.  1988.  A  comparison  of  sampling  schemes  used  in  generating 
error  matrices  for  assessing  the  accuracy  of  maps  generated  from  remotely 
sensed  data.  Photogrammetric  Engineering  and  Remote  Sensing,  54:  593-600. 


525 


Crist,  E.P.  and  R.C.  Cicone.  1984.  A  physically-based  transformation  of 
Thematic  Mapper  data  -  The  TM  Tasseled  Cap.  IEEE  Transactions  on 
Geoscience  and  Remote  Sensing  GE-22:  256  -  263. 

Dzur,  R.S.,  M.E.  Garner,  K.G.  Smith,  and  W.F.  Limp.  In  press.  Gap  Analysis 
partnerships  for  mapping  the  vegetation  of  Arkansas.  In  J.M.  Scott,  T.H.  Tear, 
and  F.  Davis,  eds.  Gap  Analysis:  A  landscape  approach  to  biodiversity 
planning.  American  Society  for  Photogrammetry  and  Remote  Sensing, 
Bethesda,  Maryland. 

Edwards,  T.C.  Jr.,  C.G.  Homer,  S.D.  Bassett,  A.  Falconer,  R.D.  Ramsey,  and 
D.W.  Wight.  1995.  Utah  Gap  Analysis:  An  environmental  information  system. 
Final  Project  Report  95-1.  Utah  Cooperative  Fish  and  Wildlife  Research  Unit. 
Utah  State  University,  Logan,  UT  84322-5210. 

Foti,  T,  X.  Li,  M.  Blaney,  and  K.G.  Smith.  1994.  A  classification  system  for  the 
natural  vegetation  of  Arkansas.  Proceedings  Arkansas  Academy  of  Sciences, 
48:  50-62. 

Janssen,  L  L.F.,  and  F.J.M.  van  der  Wei.  1994.  Accuracy  assessment  of  satellite 
derived  land-cover  data:  A  review.  Photogrammetric  Engineering  and  Remote 
Sensing,  60:  419-426. 

Jennings,  M.D.  1993.  Natural  terrestrial  cover  classification:  assumptions  and 
definitions.  Gap  Analysis  Technical  Bulletin  2.  U.S.  Fish  and  Wildlife  Service, 
Idaho  Cooperative  Fish  and  Wildlife  Unit,  University  of  Idaho,  Moscow. 

Scott,  J.M,  F.  Davis,  B.  Csuti,  R.  Noss,  B.  Butterfield,  C.  Groves,  H.  Anderson, 
S.  Caicco,F.  D'Erchia,  T.C.  Edwards  Jr.,  J.  Ulliman,  and  R.G.  Wright.  1993. 
Gap  Analysis:  a  geographic  approach  to  protection  of  biological  diversity. 
Wildlife  Monographs  123:  1  -  41. 

Verbyla,  D.L.,  and  TO.  Hammond.  1995.  Conservative  bias  in  classification 
accuracy  assessment  due  to  pixel-by-pixel  comparison  of  classified  images  with 
reference  grids.  International  Journal  of  Remote  Sensing,  16:  581-587. 

BIOGRAPHICAL  SKETCH 

Robert  S.  Dzur  is  Project  Leader  for  the  National  Gap  Analysis  project  in 
Arkansas.  He  graduated  from  the  University  of  Arkansas  in  1991  with  a  B.A. 
double  major  in  Spanish  and  Geography  and  received  an  M.A.  in  Geography  from 
the  University  of  Arkansas  in  1994.  His  research  interests  include  applications  of 
remote  sensing/GIS  to  scale  and  spatial  dependency  of  vegetation  and  landform 
relationships. 


526 


Fuzzy  Measures  in  Multi-Criteria 

Evaluation 


J.  Ronald  Eastman1  and  Hong  Jiang2 


Abstract.  —  A  consideration  of  Multi-Criteria  Evaluation  (MCE)  in 
GIS  leads  to  the  conclusion  that  the  standardized  factors  of  MCE  be- 
long to  a  general  class  of  fuzzy  measures  and  the  more  specific  instance 
of  fuzzy  sets:  a  perspective  that  provides  a  strong  theoretical  basis  for 
the  standardization  of  factors  and  their  subsequent  aggregation.  In  this 
context,  a  new  aggregation  operator  that  permits  continuous  variation  of 
ANDORness  and  tradeoff  is  discussed:  the  Ordered  Weighted  Average. 


INTRODUCTION 

Multi-Criteria  Evaluation  (MCE)  is  perhaps  the  most  fundamental  of  decision 
support  operations  in  GIS.  MCE  is  most  commonly  achieved  by  one  of  two  pro- 
cedures. The  first  involves  boolean  overlay  whereby  all  criteria  are  reduced  to 
logical  statements  of  suitability  and  then  combined  by  means  of  one  or  more 
logical  operators  such  as  intersection  or  union.  The  second  is  known  as  weighted 
linear  combination  wherein  continuous  criteria  {factors)  are  standardized  to  a 
common  numeric  range,  and  then  combined  by  means  of  a  weighted  average.  The 
result  is  a  continuous  mapping  of  suitability  that  may  then  be  masked  by  one  or 
more  boolean  constraints  to  accommodate  qualitative  criteria,  and  finally  thresh- 
olded  to  yield  a  final  decision  (Figure  1). 

Despite  the  very  common  use  of  these  procedures,  there  are  some  fundamental 
problems  associated  with  their  use.  First,  despite  a  casual  expectation  that  the  two 
procedures  should  yield  similar  results,  they  very  often  do  not  (Figure  1).  The 
reason  clearly  has  to  do  with  tradeoff,  but  also  with  the  logic  of  the  aggregation 
operation.  For  example,  boolean  intersection  results  in  a  very  hard  AND  ~  a  re- 
gion will  be  excluded  from  the  result  if  any  single  criterion  fails  to  exceed  its 
threshold.  Conversely,  the  boolean  union  operator  implements  a  very  liberal 
mode  of  aggregation  —  a  region  will  be  included  in  the  result  even  if  only  a  single 
criterion  meets  its  threshold.  Weighted  linear  combination  is  quite  unlike  these 
options.  Here  a  low  score  on  one  criterion  can  be  compensated  by  a  high  score  on 
another  —  a  feature  known  as  tradeoff  or  substitutability. 


Director,  The  Clark  Labs  for  Cartographic  Technology  and  Geographic  Analysis/IDRISI  Project,  and  Professor  of 
Geography,  Clark  University,  Worcester,  MA  01610,  USA. 

Research  Analyst,  The  IDRISI  Project,  Clark  University,  Worcester,  MA  01610,  USA. 


527 


Figure  1.-GIS  procedures 
for  Multi-Criteria 
Evaluation  (MCE) 

Left:  Boolean  evaluation 
for  areas  suitable  for 
industrial  development 
based  on  four  criteria: 
proxmity  to  road, 
proximity  to  the  main 
town,  slope  gradient  and 
distance  from  a  wildlife 
reserve.  The  bottom  is 
the  result  for  industrial 
allocation  produced  by 
the  intersection  operator. 

Right:  The  same  criteria 
as  continuous  factors, 
weighted  linear  combi- 
nation, and  at  the  bottom 
industrial  allocation  with 
an  area  equal  to  that  of 
the  boolean  result. 

Center:  At  the  top  is  the 
constraint  image  with 
the  area  that  precludes 
industrial  development. 
On  the  bottom  is  the 
cross-tabulation  image 
of  the  boolean  and 
weighted  linear  results. 


528 


The  second  problem  with  MCE  has  to  do  with  the  standardization  of  factors  in 
weighted  linear  combination.  The  most  common  approach  is  to  rescale  the  range 
to  a  common  numerical  basis  by  simple  linear  transformation.  However,  the  ra- 
tionale for  doing  so  is  unclear  (Voogd,  1983;  Eastman  et  al.,  1993).  Indeed,  there 
are  many  instances  where  it  would  seem  logical  to  rescale  values  within  a  more 
limited  range.  For  example,  if  proximity  to  roads  is  a  factor  in  determining  suit- 
ability for  industrial  development,  it  may  be  that  a  location  10  km  away  is  no  less 
suitable  than  one  that  is  5  km  away  —  they  are  both  simply  too  far  to  consider. 
Furthermore,  there  are  cases  where  a  non-linear  scaling  may  seem  appropriate. 
Since  the  recast  criteria  really  express  suitability,  there  are  many  cases  where  it 
would  seem  appropriate  that  criterion  scores  asymptotically  approach  the  maxi- 
mum or  minimum  suitability  level. 

A  third  problem  concerns  decision  risk.  Decision  risk  may  be  considered  as  the 
likelihood  that  the  decision  made  will  be  wrong  (Eastman,  1996).  For  both  proce- 
dures it  is  a  fairly  simple  matter  to  propagate  measurement  error  through  the  de- 
cision rule  and  subsequently  to  determine  the  risk  that  a  given  location  will  be 
assigned  to  the  wrong  set  (i.e.,  the  set  of  selected  alternatives  or  the  set  of  those 
not  to  be  included).  However,  the  continuous  criteria  of  weighted  linear  combi- 
nation would  appear  to  express  a  further  uncertainty  that  is  not  so  easily  accom- 
modated. The  standardized  factors  of  weighted  linear  combination  each  express  a 
perspective  of  suitability  —  the  higher  the  score,  the  more  suitable.  However, 
there  is  no  real  threshold  that  can  definitively  allocate  locations  to  one  of  the  two 
sets  involved  (areas  to  be  chosen  and  areas  to  be  excluded).  How  are  these  uncer- 
tainties to  be  accommodated  in  expressions  of  decision  risk?  If  these  criteria 
really  express  uncertainties,  why  are  they  combined  through  an  averaging  proc- 
ess? It  is  the  contention  here  that  these  and  all  of  the  issues  raised  here  can  be  re- 
solved through  a  consideration  of fuzzy  measures. 

FUZZY  MEASURES 

The  term  fuzzy  measure  refers  to  any  set  function  which  is  monotonic  with  re- 
spect to  set  membership  (Dubois  and  Prade,  1982).  Notable  examples  of  fuzzy 
measures  include  probabilities,  the  beliefs  and  plausibilities  of  Dempster- Shafer 
theory,  and  the  possibilities  of  fuzzy  sets.  Interestingly,  if  we  consider  the  process 
of  standardization  in  MCE  to  be  one  of  transforming  criterion  scores  into  set 
membership  statements  (i.e.,  the  set  of  suitable  choices),  then  standardized  crite- 
ria are  fuzzy  measures. 

A  common  trait  of  fuzzy  measures  is  that  they  follow  DeMorgan's  Law  in  the 
construction  of  the  intersection  and  union  operators  (Bonissone  and  Decker, 
1986).  DeMorgan's  Law  establishes  a  triangular  relationship  between  the  inter- 
section, union  and  negation  operators  such  that: 


529 


T(a,b)  =  ~S(~a,~b) 


where  T  -  Intersection  (AND)         =  T-Norm 

and    S  =  Union  (OR)  =  T-CoNorm 

and    ~  =         Negation  (NOT) 

The  intersection  operators  in  this  context  are  known  as  triangular  norms,  or  sim- 
ply T-Norms,  while  the  union  operators  are  known  as  triangular  co-norms,  or 
T-CoNorms. 


T-NORMS/T-CONORMS  AND  THE  AVERAGING  OPERATOR 

A  T-Norm  can  be  defined  as  (Yager,  1988): 

a  mapping  T:  [0,1]  *  [0,1]  ->  [0,1]  such  that : 

T(a,b)  =  T(b,a)  (commutative) 
T(a,b)  >=  T(c,d)  if  a  >=  c  and  b  >=  d  (monotonic) 
T(a,T(b,c))  =  T(T(a,b),c)  (associative) 
T(l,a)  =  a 

Some  examples  of  T-norms  include: 

min(a,b)  (the  intersection  operator  of  fuzzy  sets) 

a  *  b  (the  intersection  operator  of  probabilities) 

1  -  min(  1 ,(( 1  -a)Ap  +  ( 1  -b)Ap  ) A(  1  /p))  (for  p>=  1 ) 

max(0,a+b-l) 

Conversely,  a  T-CoNorm  is  defined  as: 

a  mapping  S:  [0,1]  *  [0,1]  ->  [0,1]  such  that : 

S(a,b)  =  T(b,a)  (commutative) 
S(a,b)  >=  T(c,d)  if  a  >=  c  and  b  >=  d  (monotonic) 
S(a,T(b,c))  =  T(T(a,b),c)  (associative) 
S(0,a)  =  a 

Some  examples  of  T-CoNorms  include: 

max(a,b)  (the  union  operator  of  fuzzy  sets) 

a  +  b  -  a*b  (the  union  operator  of  probabilities) 

min(  1  ,(aAp  +  bAp  )A(  1  /p))  (for  p>=  1 ) 

min(l,a+b) 

Interestingly,  while  the  intersection  and  union  operators  of  boolean  overlay  repre- 
sent a  T-Norm/T-CoNorm  pair,  the  averaging  operator  of  weighted  linear  combi- 
nation cannot  because  it  lacks  the  property  of  associativity  (Bonissone  and 


530 


Decker,  1986).  Rather,  it  has  been  determined  (Bonissone  and  Decker,  1986)  that 
the  averaging  operator  falls  mid-way  between  the  extreme  cases  of  the  T-Norm 
of  fuzzy  sets  (the  minimum  operator)  and  its  corresponding  T-CoNorm  (the 
maximum  operator)  in  the  sense  that  it  is  neither  an  AND  nor  an  OR  operator, 
but  rather,  one  that  lies  halfway  in  between  ~  in  essence,  a  perfect  ANDOR  op- 
erator. This  quality  has  interested  some  in  the  Decision  Science  field  (e.g.,  Yager, 
1988)  because  of  the  recognition  that  in  human  perception  of  decision  logics,  it  is 
not  uncommon  to  wish  to  combine  criteria  with  something  less  extreme  than  the 
hard  operations  of  union  or  intersection.  What  is  particularly  interesting  here, 
however,  is  that  the  averaging  operator  falls  along  a  continuum  of  ANDORness 
between  the  intersection  and  union  operators  of  fuzzy  sets.  Can  the  aggregation 
process  of  weighted  linear  combination  be  seen  as  fuzzy  set  membership  calcula- 
tion? Given  that  it  has  already  been  determined  that  standardized  criteria  qualify 
as  fuzzy  measures,  and  that  fuzzy  sets  are  fuzzy  measures,  the  concept  is 
plausible. 

WEIGHTED  LINEAR  COMBINATION  AND  FUZZY  SETS 

There  are  a  variety  of  reasons  why  consideration  of  the  weighted  linear  combi- 
nation process  as  a  fuzzy  set  membership  aggregation  operator  is  highly  appeal- 
ing. First,  it  provides  a  very  strong  logic  for  the  process  of  standardization.  In  this 
context,  the  process  of  standardizing  a  criterion  can  be  seen  as  one  of  recasting 
values  into  a  statement  of  set  membership  —  the  degree  of  membership  in  the  fi- 
nal decision  set.  Second,  this  clearly  opens  the  way  for  a  broader  family  of  set 
membership  functions  than  that  of  linear  rescaling.  For  example,  the  commonly 
used  sigmoidal  function  provides  a  simple  logic  for  cases  where  a  function  is  re- 
quired that  is  asymptotic  to  0  and  1.  Third,  the  logic  of  fuzzy  sets  bridges  a  major 
gap  between  that  of  boolean  overlay  and  that  of  the  weighted  linear  combination. 
Boolean  overlay  is  very  clearly  a  classical  set  problem.  By  considering  the  proc- 
ess as  one  of  fuzzy  sets,  the  process  of  weighted  linear  combination  can  also  be 
seen  as  a  set  problem.  Fourth,  in  cases  where  set  membership  approaches  cer- 
tainty, the  results  from  fuzzy  sets  will  be  identical  to  those  of  boolean  overlay. 

If  the  problem  can  indeed  be  recast  into  the  framework  of  fuzzy  sets,  is  it  rea- 
sonable that  the  minimum  and  maximum  operators  should  function  as  the  inter- 
section and  union  operators?  As  it  happens,  both  are  well  known  in  human 
experience.  The  minimum  operator  is  commonly  used  in  GIS  applications,  and 
represents  a  form  of  limiting  factor  analysis.  Here  the  intent  is  one  of  risk  aver- 
sion, by  characterizing  the  suitability  of  a  location  in  terms  of  its  worst  quality. 
The  maximum  operator  is  the  opposite,  and  can  thus  be  thought  of  as  a  very  opti- 
mistic aggregation  operator  —  an  area  will  be  suitable  to  the  extent  of  its  best 
quality.  That  the  averaging  operator  should  fall  half  way  between  these  extremes 
is  not  surprising. 


531 


the  MIN  and  MAX  operators 
Figure  2.-OWA  results  with  different  order  weights 


532 


THE  ORDERED  WEIGHTED  AVERAGE 


With  the  recognition  that  the  averaging  function  represents  an  ANDOR  opera- 
tor midway  between  the  fuzzy  intersection  and  union  operators,  and  that  it  also 
provides  full  tradeoff  between  criteria,  the  question  arises  of  the  extent  to  which 
these  two  qualities  might  be  varied  independently.  Looking  to  the  field  of  Deci- 
sion Science,  a  very  interesting  approach  is  that  of  Yager  (1988).  Yager  has  pro- 
posed a  variant  of  the  averaging  operator,  called  an  Ordered  Weighted  Average, 
that  can  achieve  continuous  control  over  the  degree  of  ANDORness  of  the  opera- 
tor and  the  degree  of  tradeoff  between  criteria.  In  his  implementation,  the  criteria 
are  weighted  on  the  basis  of  their  rank  order  rather  than  their  inherent  qualities. 
Thus,  for  example,  we  might  decide  to  apply  weights  of  0.5,  0.3,  0.2  to  weight  a 
set  of  factors  A,  B  and  C  based  on  their  rank  order.  Thus  if  at  one  location  the 
criteria  are  ranked  BAC  (from  lowest  to  highest),  the  weighted  combination 
would  be  0.5 *B  +  0.3*  A  +  0.2  *C.  However,  if  at  another  location  the  factors  are 
ranked  CBA,  the  weighted  combination  would  be  0.5 *C  +  0.3 *B  +  0.2  *A.  In  our 
implementation  of  Yager's  concept,  we  have  retained  the  concept  of  weights  that 
apply  to  specific  factors,  yielding  two  sets  of  weights  —  criterion  weights  that  ap- 
ply to  specific  criteria  and  order  weights  that  apply  to  the  ranked  criteria,  after 
application  of  the  criterion  weights. 

The  interesting  feature  of  the  Ordered  Weighted  Average  is  that  it  is  possible  to 
control  the  degree  of  ANDORness  and  tradeoff  between  factors  within  limits.  For 
example,  using  order  weights  of  [1  0  0]  yields  the  minimum  operator  of  fuzzy 
sets,  with  full  ANDness  and  no  tradeoff.  Using  order  weights  of  [0  0  1]  yields  the 
maximum  operator  of  fuzzy  sets  with  full  ORness  and  no  tradeoff.  Using  weights 
of  [0.33  0.33  0.33]  yields  the  traditional  averaging  operator  of  MCE  with  inter- 
mediate ANDness  and  ORness,  and  full  tradeoff.  Tradeoff  is  thus  controlled  by 
the  degree  of  dispersion  in  the  weights  while  ANDness  or  ORness  is  governed  by 
the  amount  of  skew.  Thus,  for  example,  order  weights  of  [0  1  0]  would  yield  an 
operator  with  intermediate  ANDness  and  ORness,  but  no  tradeoff,  while  the 
original  example  with  order  weights  of  [0.5  0.3  0.2]  would  yield  an  operator  with 
substantial  tradeoff  and  a  moderate  degree  of  ANDness. 

We  have  recently  implemented  the  Ordered  Weighted  Average  as  a  new  mod- 
ule (called  OWA)  to  be  introduced  with  the  next  release  of  the  IDRISI  GIS  soft- 
ware system.  Figure  2  illustrates  the  output  from  this  module  using  varying  order 
weights.  The  results  are  strikingly  different,  pointing  out  the  strong  dependency 
between  the  manner  in  which  the  decision  rule  is  constructed  and  the  character  of 
the  result. 

In  use,  the  OWA  module  can  be  used  in  a  single  operation,  or  as  part  of  a  hy- 
brid design.  For  example,  if  one  had  five  factors,  where  it  was  desired  that  three 
should  trade  off,  but  that  two  should  not,  the  problem  can  be  solved  in  two 


533 


stages.  In  the  first  stage,  OWA  would  be  run  using  order  weights  with  the  desired 
degree  of  tradeoff,  followed  by  a  second  run  without  tradeoff  in  which  the  result 
of  the  first  run  was  included  as  an  additional  factor.  Criterion  weights  in  the  two 
runs  can  easily  be  adjusted  to  maintain  the  original  weights  of  all  factors. 

CONCLUSIONS 

It  has  been  demonstrated  here  that  the  standardized  criteria  of  MCE  belong  to 
the  general  class  of  fuzzy  measures,  and  it  has  been  further  proposed  that  they  be- 
long to  the  more  specific  instance  of  fuzzy  sets.  The  proposal  is  appealing  in  that 
it  provides  a  consistent  theoretical  link  between  the  MCE  logics  of  boolean  over- 
lay and  weighted  linear  combination:  the  former  representing  set  membership  re- 
lations between  crisp  sets  and  the  latter  between  fuzzy  sets.  It  also  provides  a 
theoretical  basis  for  the  standardization  of  criteria:  the  process  is  seen  as  a  recast- 
ing of  criterion  values  into  set  membership  statements.  Furthermore,  the  simple 
monotonicity  requirement  of  fuzzy  measures  permits  a  much  richer  range  of  stan- 
dardization functions  than  that  of  linear  rescaling,  including  sigmoidal  and  j- 
shaped  forms.  However,  perhaps  the  most  appealing  outcome  of  this  perspective 
is  the  manner  in  which  it  opens  up  the  possibilities  for  aggregation  of  criteria.  To 
this  end,  the  Ordered  Weighted  Average  was  demonstrated  as  an  extension  to  the 
traditional  aggregation  operator  that  allows  continuous  control  over  the  degree  of 
ANDORness  and  tradeoff  between  factors. 

REFERENCES 

Bonissone,  P.P.  and  Decker,  K.,  (1986)  "Selecting  uncertainty  calculi  and  granularity: 
an  experiment  in  trading-off  precision  and  complexity.  In  L.N.Kanal  and  J.F. 
Lemmer  eds.,  Uncertainty  in  Artificial  Intelligence  (North-Holland:  Elsevier 
Science). 

Dubois,  D.,  and  Prade,  H.,  (1982)  "A  Class  of  Fuzzy  Measures  based  on  Triangular 
Norms",  International  Journal  of  General  Systems,  8,  43-61. 

Eastman,  J.R.,  (1996)  "Uncertainty  and  Decision  Risk  in  Multi-Criteria  Evaluation:  Im- 
plications for  GIS  Software  Design",  Proceedings,  UN  University  International 
Institute  for  Software  Technology  Expert  Group  Workshop  on  Software  Tech- 
nology for  Agenda' 21  :  Decision  Support  Systems,  Feb.26-Mar.8,  [in  press]. 

Eastman,  J.R.,  Kyem,  P.A.K.,  Toledano,  J.,  and  Jin,  W.,  (1993)  "GIS  and  Decision 
Making",  Explorations  in  Geographic  Information  System  Technology,  4,  (Ge- 
neva: UNITAR). 

Voogd,  H.,  (1983)  Multicriteria  Evaluation  for  Urban  and  Regional  Planning,  (London: 
Pion). 

Yager,  R.  (1988)  "On  Ordered  Weighted  Averaging  aggregation  operators  in  multicrite- 
ria decision  making",  IEEE  Transactions  on  Systems,  Man,  and  Cybernetics. 
8(1):  183-190. 


534 


On  Roles  and  Goals 
for  Map  Accuracy  Assessment: 
A  Remote  Sensing  Perspective 


Curtis  E.  Woodcock  1 


Abstract.  -  Expanded  roles  and  goals  for  accuracy  assessment  are  pos- 
sible based  on  changes  in  the  assumptions  regarding  the  properties  of 
classes  in  thematic  maps.  Existing  views  of  thematic  maps  use  classi- 
cal set  theory,  which  assumes  each  location  in  the  map  can  be  assigned 
unambiguously  to  a  single  class.  Similarly,  classes  are  assumed 
exhaustive  and  mutually  exclusive.  The  substitution  of  fuzzy  set 
theory  for  classical  set  theory  relaxes  these  assumption  and  affords  new 
opportunities  and  challenges.  The  main  concept  required  for  use  of 
fuzzy  set  theory  is  acknowledgement  of  varying  levels  of  class 
membership  for  individual  map  units.  Three  areas  where  improve- 
ments are  possible  using  accuracy  assessment  based  on  fuzzy  sets  are: 
1)  providing  additional  information  on  map  accuracy,  including  infor- 
mation on  magnitudes  of  errors,  and  alternative  methods  of  estimating 
the  frequency  of  errors;  2)  expanding  the  kinds  of  queries  that  are  pos- 
sible with  regard  to  area  estimation,  including  estimates  of  areas  of 
map  classes  as  a  function  of  class  membership  levels;  and  3)  improved 
landscape  characterization  by  explicit  acknowledgement  of  the  hetero- 
geneity within  remote  sensing  pixels  or  map  polygons.  Associated 
with  these  expanded  roles  and  goals  for  accuracy  assessment  is  a  need 
for  more  research  and  development. 


INTRODUCTION 


For  a  long  time  thematic  maps  have  been  an  important  method  for  display 
of  spatial  data.  Their  importance  continues,  and  possibly  has  been  enhanced  by 
the  explosive  growth  in  the  use  of  geographic  information  systems  (GIS). 
There  is  a  clear  consensus  that  accuracy  assessment  of  thematic  maps  is  an 
important  aspect  of  both  map  production  and  use.  The  importance  of  accuracy 
assessment  has  become  particularly  apparent  due  to  the  use  of  remote  sensing 
and  computer-based  methods  for  production  of  thematic  maps.    All  maps 


Associate  Professor  of  Geography,  Boston  University  Boston,  MA. 


535 


contain  errors,  but  the  use  of  more  automated  methods  has  brought  the  issue  of 
map  quality  and  accuracy  assessment  to  the  fore. 

Accuracy  assessment  has  served  two  primary  communities:  the  users  of 
maps  and  the  producers  of  maps.  From  the  production  perspective,  accuracy 
assessment  is  a  critical  dimension  of  the  process  of  evaluating  alternative  map- 
ping methodologies  and  sensors.  Map  accuracy  assessment  has  provided  the 
metric  for  comparisons  in  a  field  where  comparisons  are  frequent.  From  a 
user's  perspective,  accuracy  assessment  provides  important  information  on  the 
quality  of  the  maps,  which  in  turn  can  significantly  influence  the  manner  in 
which  maps  can  be  used.  While  the  motivation  for  users  is  typically  an  indica- 
tion of  the  reliability  of  the  map,  the  dominant  concern  remains  the  same  —  the 
overall  accuracy  of  the  map. 

The  purpose  of  this  paper  is  to  argue  for  expanded  roles  and  goals  for 
accuracy  assessment,  particularly  from  the  perspective  of  map  production. 
These  more  ambitious  roles  and  goals  of  accuracy  assessment  are  made  possi- 
ble by  a  change  in  the  fundamental  assumptions  about  the  nature  of  the  classes 
in  thematic  maps.  Conventionally,  the  classes  in  thematic  maps  are  assumed  to 
follow  classical  set  theory,  in  which  each  location  in  the  landscape  is  assumed 
to  belong  to  a  single  map  class.  Additionally,  the  map  categories  are  assumed 
to  be  mutually  exclusive  and  exhaustive.  Sets,  or  map  classes  in  this  case,  that 
meet  these  conditions  are  often  termed  crisp  sets.  The  continuum  of  variation 
found  in  landscapes  cannot  be  captured  by  classes  using  the  assumptions  of 
crisp  sets.  The  substitution  of  fuzzy  set  theory  for  classical  set  theory  relaxes 
assumptions  regarding  the  mutually  exclusive  nature  of  map  classes  and  affords 
new  opportunities  and  challenges.  I  suggest  three  initial  categories  of  issues 
arising  from  the  substitution  of  fuzzy  set  theory.  This  list  is  not  intended  to  be 
exhaustive,  but  rather  a  start.  Each  of  these  has  received  some  initial  research 
attention,  but  more  is  clearly  needed  and  warranted. 

Background  on  Fuzzy  Set  Theory 

The  concept  of  a  fuzzy  set  was  introduced  by  Zadeh  (1965)  to  describe 
imprecision  that  is  characteristic  of  much  of  human  reasoning.  Fuzzy  sets  pro- 
vide a  quantitative  approach  for  dealing  with  vagueness  in  complex  systems.  A 
primary  difference  between  fuzzy  set  theory  and  classical  set  theory  concerns 
membership  functions.  In  classical  set  theory,  each  object  or  element  is  either 
a  member  of  a  set  or  it  isn't.  Fuzzy  set  theory  allows  for  grades  of  member- 
ship, providing  considerable  flexibility  beyond  that  available  using  classical  set 
theory.  For  thematic  maps,  fuzzy  sets  afford  the  opportunity  to  acknowledge 
ambiguity  about  the  class  membership  of  locations  in  the  map.  Sites  with  pro- 
perties near  the  definitional  boundaries  between  classes  can  be  characterized 
accordingly. 


536 


ROLES  AND  GOALS  FOR  ACCURACY  ASSESSMENT 


Issue  1:  Enhanced  Accuracy  Information 

Gopal  and  Woodcock  (1994)  have  provided  an  initial  attempt  at  accuracy 
assessment  methods  for  thematic  maps  based  on  fuzzy  set  theory.  These 
methods  expand  the  kind  of  information  resulting  from  map  accuracy  assess- 
ment. The  basic  approach  to  accuracy  assessment  is  similar  to  conventional 
methods  with  respect  to  sample  selection  and  the  need  for  ground  truth  (Con- 
galton,  1991).  The  primary  difference  in  data  collection  is  the  requirement  for 
a  fuzzy  membership  value  for  each  possible  map  class  for  each  sample  site.  A 
series  of  fuzzy  operators  are  used  to  evaluate  the  data  collected  at  accuracy 
assessment  sites.  Two  operators,  MAX  and  RIGHT  remain  concerned  with  the 
primary  question  of  overall  map  accuracy.  However,  two  independent  meas- 
ures of  map  accuracy  result,  with  MAX  providing  a  more  conservative  estimate 
of  map  accuracy  than  RIGHT.  The  difference  in  the  results  for  these  two 
operators,  combined  with  the  results  from  a  DIFFERENCE  operator  provide 
indications  of  the  magnitude  of  the  errors  in  the  map.  Conventionally,  fre- 
quency of  errors  has  been  the  criterion  determining  overall  map  accuracy. 
Given  the  ability  to  measure  the  magnitude  of  errors,  it  would  be  beneficial  to 
incorporate  this  information  into  measures  of  overall  map  accuracy. 

The  nature  of  the  errors  in  the  map,  or  an  analysis  of  the  confusion 
between  classes  proceeds  in  a  similar  fashion  to  conventional  confusion 
matrices  using  a  CONFUSION  operator.  One  added  dimension  associated  with 
the  use  of  fuzzy  sets  is  the  AMBIGUITY  operator,  which  tracks  the  frequency 
of  ties  between  classes.  The  results  from  the  AMBIGUITY  operator  are 
arranged  in  a  matrix  similar  to  the  CONFUSION  operator,  and  are  particularly 
interesting  if  they  are  asymmetric.  In  this  situation  it  is  possible  to  infer  the 
distribution  in  the  map  of  areas  receiving  equal  fuzzy  membership  values  for 
two  different  classes. 

Overall,  the  use  of  fuzzy  operators  provides  more  information  about  the 
accuracy  of  a  map  than  is  possible  using  conventional  methods.  The  use  of 
both  the  MAX  and  RIGHT  operators  increases  the  information  about  the  fre- 
quency of  errors,  and  the  overall  accuracy  of  the  map.  Explicit  consideration 
of  the  magnitude  of  errors  is  also  an  important  development.  However,  the 
work  of  Gopal  and  Woodcock  should  be  considered  only  a  start.  The  relaxa- 
tion of  the  limitations  required  by  classical  set  theory  provides  many  opportuni- 
ties for  improved  accuracy  assessment  which  remain  to  be  explored. 

Issue  2:  An  Expanded  View  of  Area  Estimation 

The  use  of  fuzzy  sets  and  the  concept  of  levels  of  class  membership 
expand  significantly  the  kinds  of  queries  possible  regarding  the  areal  extent  of 
classes.  Data  from  conventional  map  accuracy  assessments  and  marginal  map 
totals  have  frequently  been  used  to  provide  improved  estimates  of  the  true  area 
of  map  classes  (Card,  1982;  Hay,  1988;  Jupp,  1989).  However,  these 
approaches  are  limited  to  the  constraint  imposed  by  classical  set  theory  that  the 
sums  of  the  areas  of  all  map  classes  must  be  unity.  Once  the  idea  that  various 


537 


grades  of  class  membership  exist,  it  is  possible  to  pose  questions  regarding  the 
areal  extent  of  classes  as  a  function  of  level  of  class  membership.  When 
viewed  from  this  perspective,  the  sums  of  the  areas  for  all  classes  need  not  sum 
to  unity.  In  fact,  values  below  unity  would  be  expected  for  strict  levels  of 
class  membership,  and  values  above  unity  for  more  lenient  levels  of  class 
membership. 

Card's  (1982)  methods  for  area  estimation  have  been  adapted  for  use  in 
estimating  the  area  in  maps  having  various  levels  of  class  membership.  The 
basis  for  Card's  approach  is  to  assign  the  area  in  the  map  represented  by  each 
sample  (which  is  a  function  of  the  number  of  samples  in  the  map  class,  and  the 
map  area  of  the  map  class)  to  the  true  class  of  the  sample  site.  By  retaining 
the  column  totals  (map  classes)  for  sample  sites  from  the  confusion  matrix  and 
the  marginal  map  proportions  and  substituting  a  confusion  matrix  that  includes 
only  sample  sites  exhibiting  specified  levels  of  class  membership,  it  is  possible 
to  estimate  the  true  area  of  each  class  for  specific  levels  of  class  membership 
(Woodcock  and  Gopal,  199?).  The  idea  that  the  sums  of  the  areas  of  classes 
might  not  equal  unity  is  initially  unsettling.  However,  when  viewed  from  out- 
side the  normal  questions  of  area  estimation,  it  is  easier  to  understand.  For 
example,  one  way  to  view  the  problem  is  more  like  the  general  question  of 
which  locations  in  a  map  meet  a  certain  set  of  conditions.  This  kind  of  prob- 
lem is  now  common  in  GIS,  and  expectations  for  the  results  of  several  such 
queries  to  sum  to  unity  rarely  exist. 

The  ability  to  estimate  the  areas  of  classes  as  a  function  of  levels  of  class 
membership  should  improve  the  utility  of  maps.  An  example  from  vegetation 
mapping  helps  illustrate  this  point.  Currently,  users  are  limited  to  questions 
regarding  the  areas  dominated  by  a  particular  vegetation  type.  However,  mixes 
of  vegetation  types  are  common.  Using  the  methods  mentioned  above,  it 
would  now  be  possible  to  ask  questions  regarding  the  areal  extent  of  locations 
where  a  particular  vegetation  type  is  at  least  present,  if  not  dominant.  These 
kinds  of  queries  are  particularly  important  in  wildlife  habitat  studies,  where  the 
presence  of  a  particular  species  of  tree  may  be  important  for  nesting  purposes. 

This  direction  of  research  is  obviously  new,  and  lacks  the  benefits  of  peer 
review.  However,  the  possibility  for  improving  the  information  that  can  be 
provided  to  map  users  is  considerable. 

Issue  3:  Improved  Landscape  Characterization 

One  trend  in  remote  sensing  has  been  toward  providing  maps  at  regional, 
continental,  and  global  scales.  To  accommodate  mapping  areas  of  these  sizes, 
increasing  use  is  being  made  of  satellite  imagery  with  coarse  spatial  resolu- 
tions, such  as  provided  by  the  NOAA  AVHRR  (see  for  example,  DeFries  and 
Townshend,  1994).  One  issue  arising  in  the  use  of  coarse  spatial  resolution 
imagery  in  making  thematic  maps  concerns  the  internal  homogeneity  of  indivi- 
dual pixels.  Portraying  large  areas  as  a  single  class  provides  an  imprecise  view 
of  landscapes.  Prior  research  has  indicated  significant  errors  in  the  estimation 
of  map  proportions  can  result  from  the  use  of  large  pixel  sizes,  including  the 
overestimation  of  the  area  of  larger  classes  and  the  disappearance  of  smaller 


538 


classes  (Turner  et  al.,  1989;  Moody  and  Woodcock,  1994). 

One  way  to  view  this  problem  concerns  the  ability  of  classes  based  on 
crisp  sets  to  characterize  landscapes  at  these  scales.  The  use  of  fuzzy  sets 
allows  for  explicit  acknowledgement  of  internal  heterogeneity  for  large  pixels. 
One  community  within  remote  sensing  assumes  all  pixels  are  mixtures,  and 
employs  methods  designed  to  estimate  the  proportions  of  constituent  com- 
ponents (Adams  et  al.,  1986;  Roberts,  et  al.,  1993;  Defries  et  al.,  1995).  The 
use  of  fuzzy  set  representations  for  classes  in  thematic  maps  may  provide  a 
bridge  between  this  mixture  modeling  community  and  the  production  of 
thematic  maps.  The  key  element  that  fuzzy  sets  can  provide  in  this  context  is  a 
method  of  translating  various  component  mixtures  into  levels  of  class  member- 
ship. 

DISCUSSION  AND  CONCLUSIONS 

This  paper  calls  for  a  broader  role  for  map  accuracy  assessment  based  on 
the  more  realistic  underlying  assumptions  of  fuzzy  set  theory.  The  expanded 
opportunities  carry  the  requirement  of  an  extensive  supporting  research  agenda. 
For  example,  adaptation  of  some  of  the  recent  developments  for  conventional 
accuracy  assessment  for  fuzzy  sets  would  be  beneficial  (Foody,  1992;  Ma  and 
Redmond,  1995).  While  a  substantial  research  agenda  is  needed,  it  should  be 
noted  that  the  general  steps  required  for  accuracy  assessment  based  on  fuzzy 
sets  are  quite  similar  to  conventional  methods. 

The  goals  of  accuracy  assessment  can  be  expanded  beyond  consideration 
of  the  nature  and  frequency  of  errors  in  maps  to  include  assessment  of  the 
magnitude  of  those  errors  and  alternative  measures  of  their  frequency.  Estima- 
tion of  the  area  of  map  classes  as  a  function  of  class  membership  levels 
significantly  expands  the  kinds  of  queries  that  are  possible,  and  hence  improves 
the  utility  of  maps.  Additionally,  the  use  of  fuzzy  sets  in  thematic  maps  can 
improve  landscape  characterization  by  allowing  explicit  acknowledgement  of 
heterogeneity  within  map  units. 

REFERENCES 

Adams,  J.  B,  Smith,  M.O.,  and  P.E.  Johnson,  1986.  Spectral  Mixture 
Modeling:  A  New  Analysis  of  Rock  and  Soil  Types  at  the  Viking  Lander 
1  Site,  Journal  of  Geophysical  Research,  91  (B8):8098-8 1 12. 

Card,  D.H.,  1982.  Using  known  map  category  marginal  frequencies  to  improve 
estimates  of  thematic  map  accuracy,  Photo  gramme  trie  Engineering  and 
Remote  Sensing,  48(3):43 1-439. 

Congalton,  R.G.,  1991.  A  review  of  assessing  the  accuracy  of  classifications  of 
remotely  sensed  data,  Remote  Sensing  of  Environment,  37:35-46. 

Defries,  R.S.,  and  J.R.G.  Townshend,  1994.  NDVI-derived  land  cover 
classifications  at  a  global  scale,  International  Journal  of  Remote  Sensing, 
15(17):3567-3586. 


539 


DeFries,  R.S.,  Field,  C.B.,  Fung,  I,  Justice,  CO.,  Los,  S.,  Matson,  P.A., 
Matthews,  P.A.,  Mooney,  H.A.,  Potter,  C.S.,  Prentice,  K.,  Sellers,  P.J., 
Townshend,  J.R.G.,  Tucker,  CJ.,  Ustin,  S.L.,  and  P.M.  Vitousek,  1995. 
Mapping  the  land  surface  for  global  atmosphere-biosphere  models: 
Toward  continuous  distributions  of  vegetation's  functional  properties, 
Journal  of  Geophysical  Research,  100(D10):20,867-20,882. 

Foody,  G.M.,  1992.  On  the  compensation  for  chance  agreement  in  image 
classification  accuracy  assessment,  Photo grammetric  Engineering  and 
Remote  Sensing,  58(10):  1459-1460. 

Gopal,  S.,  and  C.E.  Woodcock,  1994.  Theory  and  Methods  for  Accuracy 
Assessment  of  Thematic  Maps  Using  Fuzzy  Sets,  Photo  grammetric 
Engineering  and  Remote  Sensing,  60(2):  181-188. 

Gopal,  S.,  C.E.  Woodcock,  and  G.  Unis,  199?.  Labeling  Map  Polygons  Using 
Remote  Sensing:  Calibrating  Decision  Rules,  submitted  to  International 
Journal  of  Geographical  Information:  Analysis,  Theory  and  Decision, 
November,  1994,  revised  August,  1995,  revised  again  November  1995,  in 
press. 

Hay,  A.M.,  1988.  The  derivation  of  global  estimates  from  a  confusion  matrix, 
International  Journal  of  Remote  Sensing,  9:1395-1398. 

Jupp,  D.L.B.,  1989.  The  stability  of  global  estimates  from  confusion  matrices, 
International  Journal  of  Remote  Sensing,  10(9):  1563-1569. 

Ma,  Z.  and  R.L.  Redmond,  1995.  Tau  coefficients  for  accuracy  assessment  of 
classification  of  remote  sensing  data,  1995,  Photo  grammetric  Engineering 
and  Remote  Sensing,  61(4):435-439. 

Moody,  A.,  and  C.E.  Woodcock,  1994.  Scale-Dependent  Errors  in  the 
Estimation  of  Land-Cover  Proportions  ~  Implications  for  Global  Land- 
Cover  Datasets,  Photo  grammetric  Engineering  and  Remote  Sensing, 
60(5):585-594. 

Moody,  A,  Gopal,  S.  and  Strahler,  A.,  1995.  Sensitivity  of  neural  network 
outputs  to  subpixel  land-cover  mixtures  in  coarse  resolution  satellite  data, 
submitted  to  Remote  Sensing  of  Environment,  July  1994,  revised, 
September,  1995. 

Roberts,  D.  A.,  M.  O.  Smith,  and  J.  B.  Adams,  1993.  Green  Vegetation, 
Nonphotosynthetic  Vegetation,  and  Soils  in  AVIRIS  Data,  Remote  Sensing 
of  Environment,  44:222-269. 

Turner,  M.,  O'Neill,  R.,  Gardner,  R.  and  B.  Milne,  1989.  Effects  of  changing 
spatial  scale  on  the  analysis  of  landscape  pattern,  Landscape  Ecology, 
3(3/4):153-162. 

Woodcock,  C.E.,  and  S.  Gopal,  199?.  Fuzzy  Set  Theory  and  Thematic  Maps: 
Accuracy  Assessment  and  Area  Estimation,  in  preparation  for 
International  Journal  of  GIS. 

Zadeh,  L.,  1965.  Fuzzy  Sets,  Information  and  Control,  8:338-353. 


540 


Integration  Of  Inventory  And  Field 
Data  For  Automated  Fuzzy  Accuracy 
Assessment  Of  Large  Scale  Remote- 
Sensing  Derived  Vegetation  Maps  In 
Region  5  National  Forests 

Jeff  A.  Milliken1  and  Curtis  E.  Woodcock2 


Abstract. — The  U.S. D  A.  Forest  Service  Remote  Sensing  Lab  located  in 
Sacramento,  CA  is  currently  developing  methods  to  automate  procedures 
for  determining  the  accuracy  of  forest-wide  regional  vegetation  maps 
derived  from  Landsat  Thematic  Mapper  imagery.  Procedures  are  being 
developed  to  use  both  GPS -controlled  FIA  (Forest  Inventory  Analysis) 
data  and  additional  field  data  to  assess  the  accuracy  of  these  maps. 
Methods  based  on  fuzzy  set  theory  are  being  used  in  conjunction  with  this 
data  to  determine  the  nature,  magnitude,  frequency,  and  source  of  error  in 
the  maps.  This  will  assess  the  error  for  lifeform  vegetation  labels, 
CALVEG  labels,  tree  size,  and  tree  cover.  The  key  step  under 
development  is  the  definition  of  rules  for  translating  field  measurements  of 
vegetation  properties  into  fuzzy  membership  values  for  the  various  map 
classes.  GIS  layers  are  utilized  for  developing  necessary  data  tables  for 
the  accuracy  assessment  procedure. 


INTRODUCTION 

This  paper  is  a  preliminary  report  of  methods  currently  being  developed  to  define 
and  automate  procedures  for  determining  the  accuracy  of  forest-wide  regional 
vegetation  maps  derived  from  Landsat  Thematic  Mapper  imagery  and  GIS 
modeling.. 

The  Remote  Sensing  Lab  in  Region  5  is  mapping  existing  vegetation  for  the 
National  Forests  utilizing  a  mapping  system  developed  in  conjunction  with  Boston 
University  (Woodcock  et  al.  1994  and  Miller  et  al.  1994).  These  vegetation  layers 
are  used  in  the  management  of  the  National  Forests.  For  the  proper  utilization  of 


Remote  Sen  sing  GIS  Specialist,  U.S.  Forest  Service  Region  5  Remote  Sensing  Lab,  Sacramento,  CA 

I 

Associate  Professor  of  Geography.  Boston  University 


541 


this  data,  it  is  important  to  provide  users  with  as  much  information  as  possible 
about  any  error  or  limitations  in  the  data.  All  maps  have  error  but  it  is  often 
difficult  to  obtain  accuracy  information  on  existing  data  as  conducting  an  accuracy 
assessment  on  a  data  layer  can  be  a  cost  prohibitive  process. 

The  Remote  Sensing  Lab  is  currently  developing  a  methodology  with  Boston 
University  to  assess  systematically  the  accuracy  of  these  vegetation  layers  using 
fuzzy  sets  (Woodcock  and  Gopal  1992)  and  an  independent  field  data  sample 
from  the  Forest  Inventory  and  Analysis  (FIA)  Permanent  Plot  Grid  System. 
Utilization  of  the  FIA  Grid  data  will  provide  for  a  cost  effective  means  of 
conducting  accuracy  assessment. 

ACCURACY  ASSESSMENT  METHODS 

Field  Data 

The  FIA  Permanent  Plot  Grid  consists  of  five-point  cluster  plots  sampled  on  a 
3.4  mile  grid  (U.S.  Forest  Service  1995).  Densification  of  the  grid  is  sometimes 
necessary  to  adequately  represent  under-sampled  forest  types.  In  addition  to 
collecting  the  cluster  plot  data,  inventory  crews  also  fill  out  an  accuracy 
assessment  form  for  each  point  in  the  plot.  This  form  is  intended  to  capture  "expert 
evaluation"  of  "best  and  second  best"  primary  lifeform,  CALVEG  type  (U.S. 
Forest  Service  1981),  tree  size,  and  tree  density;  as  well  as  the  same  ratings  for 
secondary  vegetation  types  where  appropriate.  This  form  may  be  modified  as 
results  from  the  accuracy  assessment  methodology  are  evaluated. 

Development  Of  Accuracy  Assessment  Fuzzy  Ratings 

Lifeform  and  CALVEG 

Map  labels  from  the  vegetation  layer  to  be  evaluated  are  added  to  the  field  data 
files  by  "overlaying"  the  locations  of  the  five-point  cluster  plots  on  the  vegetation 
layer  in  a  Geographic  Information  System.  Matrices  are  developed  to  compare  the 
map  label  assigned  to  each  cluster  point  with  the  expert  evaluations  on  the 
accuracy  assessment  form.  Fuzzy  ratings  are  assigned  to  these  unique 
combinations  based  on  descriptions  of  the  mapped  vegetation  types,  "best  and 
second  best"  expert  evaluations  from  the  accuracy  assessment  form,  and  the 
density  of  the  forest  stand  where  appropriate.  For  example,  in  a  20%  forest  stand 
with  a  shrub  under-study,  a  shrub  label  would  get  a  higher  rating  of  accuracy  than 
in  the  case  of  a  60%  (denser)  forest  stand  with  a  shrub  label.  Fuzzy  ratings  consist 
of: 

5)  Absolutely  right:  No  doubt  about  the  match.  Perfect. 
4)  Good  Answer:  Would  be  happy  to  find  this  answer  given  on  the  map. 
3)  Reasonable  or  acceptable  answer:  Maybe  not  the  best  possible  answer  but  it  is 
acceptable. 


542 


2)  Understandable  but  wrong:  Not  a  good  answer.  There  is  something  about  the 
site  that  makes  the  answer  understandable  but  there  is  clearly  a  better  answer. 

1 )  Absolutely  wrong:  This  answer  is  absolutely  unacceptable. 
(Woodcock  and  Gopal  1992). 

Tree  size  and  density 

The  Forest  Inventory  Analysis  (FIA)  software  system  allows  for  the  generation 
of  a  summary  report  by  position  for  each  cluster  plot.  This  report  gives  calculated 
values  of  mean  tree  dbh  (diameter  at  breast  height),  mean  crown  diameter,  and 
mean  canopy  cover  percent  for  predominant,  dominant  and  codominant  trees  in  the 
top-story,  and  intermediate  and  suppressed  trees  in  the  under-story  (U.S.  Forest 
Service  1995).  Using  these  calculated  figures  may  prove  to  be  a  more  unbiased 
procedure  for  assessing  the  accuracy  of  tree  size  and  density  map  labels  Both  the 
mean  and  their  standard  errors  are  used  in  assigning  fuzzy  membership  values. 

Final  Accuracy  Results 

Special  considerations  will  be  necessary  in  reporting  final  accuracy  figures  for 
the  nature,  magnitude,  frequency,  and  source  of  error  in  the  maps.  Results  will 
need  to  be  weighted  with  respect  to  such  variables  as  grid  sampling  density, 
representation  of  vegetation  types  sampled  by  the  grid,  and  the  five-point  cluster 
plot. 

CONCLUSIONS 

The  accuracy  assessment  methodology  currently  under  development  represents  a 
potentially  cost  effective  means  of  systematically  providing  important  accuracy 
information  for  large  scale  remote-sensing  derived  vegetation  maps  being  created 
in  the  Region  5  Remote  Sensing  Lab.  This  type  of  information  is  integral  to  a 
user's  ability  to  responsibly  utilize  these  maps  for  management  decisions  in  the 
National  Forests.  The  key  to  the  success  of  this  effort  is  the  ability  to  translate 
field  measurements  into  fuzzy  membership  values  for  map  classes. 

ACKNOWLEDGMENTS 

The  authors  wish  to  thank  Kevin  Casey,  John  Collins,  Kama  Kennedy,  Ralph 
Warbington,  and  Ann  Withers  for  project  assistance. 

REFERENCES 

Miller,  S.,  H.  Eng,  M.  Byrne,  J.  Milliken,  M.  Rosenberg.  1994.  Northeastern 
California  Vegetation  Mapping:  A  Joint  Agency  Effort.  Proceedings  of  the  Fifth 


543 


Forest  Service  Remote  Sensing  Applications  Conference.  Portland,  Oregon,  pp. 
115-125. 

U.S.  Forest  Service  Regional  Ecology  Group.  1981.  CALVEG:  A  Classification 

of  California  Vegetation.  U.S.  Department  of  Agriculture,  Forest  Service, 

Region  Five.  San  Francisco,  CA.  168  p. 
U.S.  Forest  Service  -  Region  5.  1995.  Forest  Inventory  and  Analysis  User's  Guide. 

U.S.  Department  of  Agriculture,  Forest  Service,  Region  Five.  San  Francisco, 

CA 

Woodcock,  C.E.  and  S.  Gopal.  1992.  Accuracy  Assessment  of  Stanislaus  Forest 
Vegetation  Map  Using  Fuzzy  Sets.  Proceedings  of  the  Fourth  Biennial  Remote 
Sensing  Applications  Conference.  Orlando,  Florida,  pp.  378-394 

Woodcock,  C.E.,  J.B.  Collins,  S.  Gopal,  V.  D.  Jakabhazy,  X.  Li,  S.  Macomber,  S. 
Ryherd,  Y.  Wu,  V.J.  Harward,  J.  Levitan,  and  R.  Warbington.  1994.  Mapping 
Forest  Vegetation  Using  Landsat  TM  Imagery  and  a  Canopy  Reflectance  Model. 
Remote  Sensing  of  Environment,  50:240-254. 

BIOGRAPHICAL  SKETCH 

Jeff  A.  Milliken  is  a  geologist  and  remote  sensing  specialist  in  the  U.S.  Forest 
Service  Remote  Sensing  Lab  located  in  Sacramento,  CA.  He  graduated  from 
Colorado  State  University  in  1975  with  a  B.S.  in  Geology  and  is  currently 
completing  his  graduate  work  in  Geography/Remote  Sensing  at  San  Francisco 
State  University. 

Curtis  E.  Woodcock  is  an  Associate  Professor  of  Geography  at  Boston 
University.  His  academic  degrees  are  from  the  University  of  California,  Santa 
Barbara,  with  interests  in  many  aspects  of  remote  sensing. 


544 


Uncertainty  of  Spatial  Metric 
Relations  in  GIS 

Xiaoyong  Chen1,  Takeshi  Doihara1  and  Mitsuru  Nasu1 

Abstract. — In  this  paper,  after  an  introduction  to  the  notations  of  metric 
topology,  a  novel  theory  of  spatial  metric  relations  between  sets  is  developed 
in  which  the  relations  are  defined  in  terms  of  the  intersections  of  the 
boundaries,  interiors  and  exteriors  of  two  dynamically  generated  sets.  Then, 
the  presented  theory  is  extended  to  quantitatively  deriving  the  spatial  metric 
relations  between  sets  in  consideration  of  conceptual  and  positional 
uncertainties  based  on  the  fuzzy  set  theory.  Finally,  some  potential 
applications  of  presented  theories  and  the  ideas  for  spatial  and  temporal 
reasoning  in  GIS  are  also  suggested.  

INTRODUCTION 

One  of  the  most  fundamental  requirements  for  GIS  is  to  modeling  and 
communicating  error  in  spatial  databases.  With  increased  research  into  error 
modeling  in  GIS  over  the  past  few  years,  there  has  a  considerable  body  of  models 
and  techniques  available  for  measurement  spatial  database  error  from  researching  to 
real  applications  [Goodchild  and  Gopal,  1989;  Vergin,  1994].  Spatial  relationships 
(such  as  distance,  direction,  ordering,  and  topology)  between  objects,  as  very  useful 
tools  for  spatial  queries,  constraints  checking  and  sorting  in  GIS,  are  strongly 
influenced  by  the  uncertainties  of  original  spatial  data.  These  practical  needs  in  GIS 
have  led  to  the  investigation  of  formal  and  sound  methods  for  driving  spatial 
relations  and  their  variations  with  uncertainties  [Chen  and  et.al.,  1995,  1996; 
Egenhofer  and  Franzosa,  1991;  Frank,  1992;  Kainz,  et.al.,  1993;  Mark,  et.  al.,1995; 
Peuquet  and  Zhang,  1987].  However,  how  to  derive  spatial  relations  between  sets 
(non-point-like)  in  consideration  of  conceptual  and  positional  uncertainties  based  on 
an  mathematically  well-defined  algebra  framework  is  still  an  open  problem  up  to 
now.  The  lack  of  this  comprehensive  theory  has  been  a  major  impediment  for 
solving  many  sophisticated  problems  in  GIS,  such  as  formally  deriving  complex 
spatial  relations  among  spatial  objects  with  multiple  representations,  and  generation 
of  the  related  standards  for  transferring  spatial  relations. 

This  paper  focuses  on  the  development  of  the  theory  and  associated  models  of 
spatial  metric  relations  between  sets  in  consideration  of  data  uncertainties.  At  first, 
based  on  the  metric  topology,  a  novel  theory  of  spatial  metric  relations  between  sets 
is  developed  in  which  the  relations  are  defined  in  terms  of  the  intersections  of  the 
boundaries,  interiors  and  exteriors  of  two  dynamically  generated  sets.  Then,  the 


AAS  Research  Institute,  Asia  Air  Survey  Co.,  LTD., 8-10,  Tamura-Cho,  Atsugi-Shi,  Kanagawa  243,  JAPAN 


545 


presented  theory  is  extended  to  quantitatively  deriving  the  spatial  metric  relations 
between  sets  in  consideration  of  conceptual  and  positional  uncertainties  based  on 
the  fuzzy  set  theory.  Finally,  some  potential  applications  of  presented  theories  and 
the  ideas  for  spatial  and  temporal  reasoning  in  GIS  area  are  also  suggested. 

FUNDAMENTAL  DEFINITIONS 

Metric  Spaces 

A  metric  space  is  a  pair  consisting  of  a  set  E  and  a  mapping  (/?,  ,p2 )  -^d(pl  ,p2 )  of 
ExE  into  R+ ,  having  the  properties:  (1)./? ,=p2<=>  d(p1>p2)=0',  (2).d(p1>p2)=d(p2,pl) 
(symmetry);  (3).  dip^p^dip^p^+dip^pj  (triangle  inequality).  The  function  d  is 
called  a  metric  and  d{px,p2)  is  called  the  distance  between  the  points  ^and  p2. 
Distance  between  points  /7.(jc.1^i:.2r_rx-.B)  in  R"  is  described  in  terms  of  the  Minkowski 
dt  -metric: 

y-J 

Conventional  Euclidean  distance  is  defined  by  the  d2  -metric.  Similarly,  the 
Manhattan  distance  defined  by  the  dl  -metric,  and  the  maximum  distance  defined  by 
the  d  -metric. 

oo 

Metric  Topological  Spaces 

Topological  spaces 

A  topological  space  is  a  pair  consisting  of  a  set  E  and  a  collection  A  of  subsets  of 
E  called  the  open  sets,  satisfying  the  three  following  properties:  (1).  every  union 
(finite  or  otherwise)  of  open  sets  is  open;  (2).  every  finite  intersection  of  open  sets  is 
open;  (3).  the  set  E  and  the  empty  set  0  are  open. 

Metric  topology 

A  metric  d  on  a  set  E  includes  a  topology  on  E,  called  metric  topology  defined  by 
d.  This  topology  is  such  that  UaE  is  an  open  set  if,  for  each  pxdj ,  there  is  an  £>o 
such  that  the  d-ba\\  of  radius  s  around  px  is  contained  U  .  A  J-ball  is  the  set  of  points 
whose  distance  from  px  in  the  metric  d  is  less  than  e  ,  i.e.  {p2^E\d(p1,p2)<e} .  Notice 
that  the  metric  topological  spaces  are  Hausdorff  and  separable. 

The  Hausdorff  Metric 

The  Hausdorff  metric  is  defined  on  the  space  K  where  each  point  is  a  non  empty 
compact  set  of^2.  If  Kx  and  K2  denote  two  non  empty  compact  set  in  R1  (or 
equivalently  two  points  in  K ),  and  B  (s)  is  the  closed  ball  with  a  radius  e ,  then  the 
quantity: 


546 


(^>) 

/[  B  J  \ 

/  (An 

I  By/ 

rtv) 

ro  o  n 

L??lJ 

0  0  !  1 

L  o  o  i  J 

ri  o  oi 

ri  o  oi 

0  10 

Lo  o  iJ 

[?::] 

[ill] 

Lo  o  i J 

fool 

L!!?J 

[i i  0 

disjoint 

inside 

contains 

equal 

meet 

covers 

coveredBy 

overlay 

Figure  1.  — The  eight  topological  relations  between  regions  in  IR2 . 


p(Kx  JC2  )=inf {e:Kx  <^K2  ®B(£),K1  c/C,  ®B(e) }  [2] 

defines  a  metric  p  on  K  ,  known  as  the  Hausdorff  metric.  From  equation  [2],  p  is  the 
radius  of  the  smallest  closed  ball  B  such  that  both  Kx  is  contained  in  the  set  K2@B(e) 
generated  by  the  morphological  dilation  [Serra,  1982]  and  K2  is  contained  in  the 
dilated  set  Kx ®B(s) .  Serra  (1982)  has  proven  that  the  Hausdorff  distance p{K^Kz) 
satisfies  all  properties  of  distance  functions.  In  particular  case,  when  Kx  and  K2 
are  reduced  to  two  points,  the  Hausdorff  distance  piK^K^  coincides  with  the 
Euclidean  distance. 


SPATIAL  METRIC  RELATIONS  BETWEEN  SETS 


Set  Intersection  Models 


9-Intersection 

For  driving  binary  topological  relations  between  sets,  Egenhofer  et  al.,  (1994) 
developed  the  9-intersection  model  based  on  the  usual  concepts  of  point-set 
topology  with  open  and  closed  sets,  in  which  the  binary  topological  relations 
between  two  objects,  Kx  and  AT, ,  in  IR1  is  based  upon  the  intersection  of  Kx 's 
interior  ( K°x  ),  boundary  (  dKx  ),and  exterior  (  K[  )  with  K2 's  interior  ( K°2 ),  boundary 
{dK2 ),  and  exterior  ( K2  ).  A  3x3  matrix 39 ,  called  the  9-intersection  as  follows: 


39^ 


K^nK;  K°xr\dK2  K°nK~2 
dKx  nK°2  dKx  ndK2  dKx  nK~2 
K:f^K:  K:r\dK  K~  nK~ 


[3] 


By  considering  the  values  empty  (0)  and  non-empty  (1)  in  equation  [3],  one  can 
distinguish  between  29=512  binary  topological  relations  in  which  only  a  small  subset 
can  be  realized  when  the  objects  of  concern  are  embedded  inlR2 .  Egenhofer  and 
Franzosa  (1991)  showed  that,  for  two  regions  with  connected  boundaries  embedded 
in  IR2 ,  the  9-intersection  distinguishes  just  8  different  relations,  i.e.  disjoint,  contains, 
inside,  equal  meet,  covers,  coveredBy,  and  overlap  (figure  1).  However,  when  we 
apply  the  9-intersection  model  to  describing  topological  relations  between  other 
types  of  spatial  objects,  the  situation  will  be  more  complicated.  According  to  the 
results  of  Mark  et.  al.  (1995),  for  two  simple  lines  33  different  spatial  relations  are 


547 


possible,  and  for  a  line  and  a  region,  19  are  possible. 


Dynamic  9-intersection 

For  integrally  deriving  different  kinds  of  spatial  relations  between  sets,  Chen  et 
al.,  (1995,  1996)  developed  the  dynamic  9-intersection  model  based  on  the  concepts 
of  the  metric  topology  with  open  and  closed  sets  and  the  morphological  dilation,  in 
which  the  general  9-intersection  of  equation  [3]  is  extended  as  follows: 

'[K^Bis^YnK;  [Kl®B(£.)]0nc>K2  [Kl©B(ei)Y  r\K'2~ 

)=  %Kx ®B(^ )^K:         )^K2  <^ )]^k;  [4] 

where  the  Kx  and  K2  are  given  two  closed  sets,  the  K.@B(et)  means  relevant 
morphological  dilation  by  the  closed  ball  B  with  radius  et ,  and  the  3^  (s, )  means 
dynamic  9-intersection  with  parameter  et  from  K.  to  K. .  Based  on  the  equation  [4], 
we  can  derive  dynamic  topological  relations  by  using  the  different  parameter  e( . 

In  particular  case,  when£(=0,  the  structure  element B(et) is  reduced  to  the 
original  point  {o},  according  to  the  algebraic  properties  of  morphological  dilation 
[Serra,  1982],  we  have  K.@B(£t)=K.®{o}=K. ,  then  the  dynamic  9-intersections 
3'iy)(£()  defined  in  equation  [4],  coincide  with  the  general  9-intersection  39  in 
equation  [3]. 


Spatial  Metric  Relations  between  Sets 

Distance  Relations  between  Sets 

According  to  the  derived  dynamic  topological  relations  by  the  dynamic  9- 
intersections  of  equation  [4]  with  different  parameter  et ,  such  as  dynamic  equal, 
dynamic  covers  (or  dynamic  corveredBy)  and  dynamic  contains  (or  dynamic 
inside),  we  can  easily  get  the  HausdorfT distance p(Kx,K2)  between  two  closed  sets 
Kx  and  K2  by  calculating  the  minimum  and  maximum  dilated  distances: 

p( Kx  ,K2  )=max{min(e. ),  min(£; )};  when 


3 


100 
010 
001 

100" 
010 
001 


000 
010 
001 

000' 
010 
001 


111 

001 
001 


100 
100 

111 


011 
001 
001 


001 
011 
001 


000 
100 

111 


000 
010 

111 


001 

111 

001 


010 
010 

111 


111 

101 
001 


110 
100 

111 


1*1 
*11 

001 

1*0' 
*10 

111 


equal 


[5] 


where  "*  "  means  either  empty  (0)  or  non-empty  (1).  The  binary  distance  relations 
derived  by  equation  [5]  are  suitable  for  different  types  of  spatial  objects,  such  as 
point-objects,  line-objects  and  region-objects,  as  well  as  combining  different  types 
of  spatial  objects  such  as  a  line  and  a  region,  a  point  and  a  line,  or  a  point  and  a 
region.  Some  examples  are  shown  in  figure  2. 


548 


Figure  2.  — The  spatial  metric  relations  between  two  planar  sets  A  and  B. 


Directional  Relations  between  Sets 

Directional  relations  between  sets  can  be  defined  by  the  Hausdorff  metric  of 
angular  bearings,  in  which  the  direction  from  one  spatial  object  to  another  is 
identical  to  that  for  metric  function  except  that  the  angular  bearing  is  computed  for 
each  ordered  pair  in  the  Cartesian  product.  The  angular  bearing  is  measured  in  the 
sense  of  navigation  bearings  (i.e.  increasing  clockwise  from  north). 

Similarly,  for  calculation  of  the  directional  relation  q>{KttKj)  between  two  non 
empty  compact  sets  Kl  and  K2  in  R2 ,  we  can  select  the  angular  bearing  set  R(at )  as 
instead  of  the  closed  ball5(e()in  equation  [3],  then  we  can  get  the  Hausdorff 
direction  (p(KitK.)  by  calculating  the  minimum  and  maximum  dilated  angles  as 
follows: 

(jp(A:i,^2)=max{min(ai),min(aj±^)}  1 
^JC,  ,*2)±*  f  Wnen 


100 

000 

1 11 

011 

001 

001 

111 

1*1 

010  . 

010 

001  , 

001 

011 

) 

111 

t 

101 

*n 

001 

001 

001 

001 

001 

001 

001 

001 

100 

000 

100 

000 

000 

010 

110 

1*0 

010 

,  010 

100 

,  100 

01  0 

010 

■• 

100 

*10 

001 

001 

111 

1 1 1 

1 1 1 

111 

1 1 1 

1 1 1 

equal  contains  covers 


549 


where  "  *  "  means  either  empty  (0)  or  non-empty  (1).  The  binary  directional  relations 
derived  by  [6]  are  also  suitable  for  different  types  of  spatial  objects  and  their 
combining  types.  Some  examples  are  shown  in  figure  2. 


UNCERTAINTY  OF  SPATIAL  METRIC  RELATIONS 


Conceptual  Uncertainty  of  Spatial  Metric  Relations 


Unlike  the  spatial  metric  relations  between  points,  the  concepts  of  distances  and 
directions  between  subsets  are  fuzzy  [Zadeh,  1965],  since  the  spatial  objects  may 
contain  many  subsets,  the  distances  and  directions  between  these  subsets  are 
difficult  to  be  represented  just  by  a  single  value.  If  we  define  the  fuzzy  memberships 
as  the  covering  percentages  of  generated  region  areas  (  or  point  numbers  and  line 
lengths  )  by  the  dynamic  intersection  of  sets  in  R2 ,  we  can  find  that  the  Hausdorff 
metric  is  just  the  special  case  with  the  fuzzy  membership  value  equal  to  one.  Based 
on  the  changes  of  the  covering  areas  of  regions  (or  point  numbers  and  line  lengths) 
from  an  empty  set  to  a  complete  set,  we  can  estimate  the  fuzzy  memberships  from 
zero  to  one,  then  we  can  quantitatively  derive  the  spatial  metric  relations  between 
subsets. 

For  reasons  of  simplicity  the  distances  and  directions  between  closed  sub- 
regions  discussed  in  this  paper  only,  related  models  for  estimation  of  conceptual 
fuzzy  membership  functions  are  defined  as  follows: 


where  A{*}  means  covered  area  sizes  by  dynamic  intersections  with  the  parameters 
A  and  0  for  distances  and  directions  separately,  the  functions  0<0>^(A)<1  and 
0<nA(9)<\  with  the  parameters  b<X<p(Kx ,K2 )  and  §<9<(p{Kx,K2)±K  are  called  the  size 
distribution  functions  [Serra,  1982].  An  example  of  conceptual  uncertainty  of  spatial 
distance  relations  between  two  spatial  regions  is  shown  in  figure  3. 


Generally,  the  locations  of  objects  in  spatial  databases  are  not  error-free,  they  may 
contain  many  kinds  of  errors,  such  as  the  errors  of  scanning,  digitizing,  selecting, 
projection,  overlaying,  and  et.al.  [Goodchild  and  et.al.,  1989].  For  describing  the 
uncertainty  in  the  positions  of  spatial  objects  (such  as  lines  and  areas),  we  can  use 
the  error  model  of  e  -band  developed  by  Chrisman  (1982),  in  which  the  positional 
uncertainty  of  a  spatial  object  K.  can  be  represented  as  K=K.©/j.(s)  ,  where  e  is 
the  buffer  distance  of  error  distribution  and  is  the  fuzzy  membership  function 
derived  by  e  .  According  to  the  equations  of  [2] -[7],  if  we  use  the  uncertain  spatial 
object  k  as  instead  of  the  general  spatial  object  K.  in  the  equations  of  [2] -[7],  we 


a>.(A)=- 


A{[KX  ®B(X)]nK2  }+A{Kx  n[K2  0£(A)] } 

A{KX)+A{K2) 
A{[K}  ®R(6)]nK2  }+A{Kx  n[K2  ®R(6±k)]  } 

A(KX)+A{K2) 


[7] 


Positional  Uncertainty  of  Spatial  Metric  Relations 


550 


<Da(?0 


Figure  3.  — The  conceptual  fuzzy  membership  function  between  subsets. 


can  drive  the  spatial  metric  relations  between  uncertain  sets  and  the  conceptual 
fuzzy  membership  functions  between  uncertain  subsets.  For  quantitatively 
estimating  the  positional  fuzzy  membership  functions  of  derived  spatial  metric 
relations  between  uncertain  subsets,  we  can  use  the  measurement  of  dynamic 
covering  uncertain  areas  as  follows: 

A{[Kx ®B(X)]nK2}+A {K  n[K,  ©ff (A)] } 


0,(A)=- 

A(Ki)+A(K2) 


[8] 


A{Kx)+A{K2) 

where  K=K-K=[Ki®h(e)]-Ki  means  the  uncertain  area  generated  by  the  related  e  - 
band,  A{*}  means  the  covered  size  of  area  or  volume  when  is  a  constant  or  a 
general  fuzzy  member  function  with  the  value  0<^(*)<1  separately.  An  example  of 
positional  uncertainty  of  spatial  directional  relations  between  two  spatial  regions  is 
shown  in  figure  4. 


CONCLUSIONS  AND  OUTLOOKS 


As  the  natural  extension  of  the  general  9-intersection,  the  dynamic  9-intersection 
based  on  metric  topology  supplied  a  general  framework  for  studying  metric  spatial 
relations  between  sets.  The  presented  integrated  theory  of  spatial  relations  between 
sets  makes  a  new  way  for  formally  deriving  spatial  relations  among  complex  spatial 
objects  with  conceptual  and  positional  uncertainties.  Even  though  the  presented 
approach  is  only  focus  on  the  applications  in  GIS  field,  the  related  results  for 
deriving  spatial  relations  between  sets  and  uncertain  sets  can  be  also  used  for  many 
other  fields,  such  as  CAD,  computer  vision,  pattern  recognition,  robot  space 
searching  and  so  on.  However,  only  the  theoretical  models  have  be  presented  in  this 


551 


paper,  a  wide  field  of  practical  application  for  management  and  analysis  of  spatial 
data  in  2-D  and  3-D  GIS  environments  has  not  been  touched.  Therefore,  the 
reported  results  must  be  verified  and  extended  in  order  to  be  used  in  different 
practical  environments. 

Two  main  directions  for  further  research  shall  be  pointed  here,  one  is  the 
applications  of  the  presented  theoretical  models  and  algorithms  in  2-D  and  3-D  GIS 
environments  for  developing  the  new  tools  of  spatial  query  and  analysis;  another  one 
is  the  extensions  of  presented  theories  and  models  for  formally  deriving  complex 
spatial  relations  among  spatial  objects  with  multiple  representations. 

ACKNOWLEDGMENTS 

The  authors  wish  especially  to  thank  Prof.  S.Murai  and  Mr.  H.Yamamoto  for  their 
considerable  helps  and  supports  of  this  research  project. 

REFERENCES 

Chen,  X.,  et.  al.,  1995.  Spatial  Relations  of  Distances  between  Arbitrary  Objects  in 
2-D/3-D  Geographic  Spaces  based  on  the  Hausdorff  Metric.  LEESMARS'95, 
Wuhan,  China.  30-41  p. 

Chen,  X.,  1996.  Spatial  Relations  between  Sets.  ISPRS  XVIII  Congress.  Vienna. 
Austria,  (accepted) 

Chrisman,  N.  R.,  1982.  Methods  of  spatial  analysis  based  on  errors  in  categorical 

maps.  Unpublished  PhD  thesis,  University  of  Bristol. 
Egenhofer,  M.,  and  Franzosa,  R.,  1991.  Point-Set  Topological  Spatial  Relations. 

IJGIS.  5(2),  161-174  p. 
Egenhofer,  M.,  and  et.  al.,  1994.  A  Critical  Comparison  of  the  4-Intersection  and  9- 

Intersection  Models  for  Spatial  Relations:  Formal  Analysis,  AUTO  CARTO  1 1 . 

1-11  p. 

Frank,  A.,  1992.  Qualitative  Spatial  Reasoning  about  Distances  and  Directions  in 
Geographic  Space.  J.  of  Visual  Languages  and  Computing.  3(4),  343-371  p. 

Goodchild,  M.,  and  Gopal,  S.,  1989.  The  Accuracy  of  Spatial  Databases.  Taylor  & 
Francis. 

Kainz,  W.,  and  et.  al.,  1993.  Modeling  Spatial  Relations  and  Operations  with 

Partially  Ordered  Sets,  IJGIS.  7(3),  215-229  p. 
Mark,  D.  J.,  et.  al.,  1995.  Toward  a  Standard  for  Spatial  Relations  in  SDTS  and 

Geographic  Information  Systems.  GIS/LIS'95,  Nashville.  Tennessee.  686-695  p. 
Peuquet,  D.  J.,  and  Zhan,  C,  1987.  An  Algorithm    to  Determine  the  Directional 

Relationship  between  Arbitrarily-Shaped  Polygons  in  the  Plane.  Pattern 

Recognition.  20(1),  65-74  p. 
Serra,  J.,  1982.  Image  Analysis  and  Mathematical  Morphology.  Academic  Press. 
Vergin,  H.,  1994.  Integration  of  Simulation  Modeling  and  Error  Propagation  for  the 

Buffer  Operation  in  GIS.  PE  &  RS.  Vol.60,  No.4,  427-435  p. 
Zadeh,  L.  A.,  1965.  Fuzzy  Sets.  Information  and  Control,  vol.  8,  338-352  p. 


552 


Plot  Collocation  Error:  Impacts  on 

Area  Estimation 


Willem  W.  S.  van  Hees  1 


Abstract-Results  of  a  study  conducted  to  examine  area  estimation  error 
caused  by  improper  collocation  of  aerial  photo  plots  and  associated 
ground  plots  are  presented.  Analyses  considered  the  complexity  of  the 
land  cover  pattern  on  the  plot,  cardinal  direction  of  collocation  error,  and 
percentage  of  area  by  land  cover  class  on  control  plots. 

During  the  1980's,  the  Forest  Inventory  and  Analysis  (FIA)-Anchorage 
unit  of  the  Pacific  Northwest  Research  Station  conducted  renewable 
resource  inventories  employing  a  three-phase-with-subsampling 
inventory  design.  There  were  three  remotely  sensed  samples  and  one 
ground  sample.  Regression  estimators  were  developed  for  several 
resource  quantities,  particularly  area  by  land  cover  class.  Results  from 
two  inventories  showed  low  correlations  between  most  covariates  on  the 
different  sampling  layers.  Statistical  consultation  indicated  inaccurate 
collocation  of  sample  plots  could  contribute  to  poor  correlations. 

This  study  focused  on  the  effects  of  improper  collocation  between  low 
altitude  aerial  photo  plots  and  ground  plots  on  estimation  of  productive 
forestland  area.  Results  indicate  there  may  be  a  north-south  directional 
bias  and  that  land  cover  complexity  as  estimated  by  fractal  dimension  is 
an  important  interaction  component.  Area  estimation  errors  ranged  from 
zero  for  homogeneous  plots  to  10.9  percent  for  plots  with  more  complex 
land  cover  patterns. 


INTRODUCTION 

During  the  1980's  the  Anchorage  Forest  Inventory  and  Analysis  (FLA)  unit  of 
the  Pacific  Northwest  Research  Station,  conducted  experimental  multiresource 
inventories  employing  a  three-phase  sampling  with  subsampling  design 
(Schreuder,  Gregoire,  Wood  1993).  Landsat  multispectral  imagery  (LS),  high- 
altitude    color    infrared    photography    (HAP),    low-altitude    color  infrared 


Research  forester,  Pacific  Northwest  Research  Station,  Anchorage,  AK. 


553 


photography  (LAP),  and  ground  (G)  plots  were  the  sampling  layers.  A  grid  of  8 
ha  plots  on  the  LS  layer  was  sampled  with  successively  more  extensive  subgrids 
on  HAP,  LAP,  and  G.  The  subgrid  of  LAP  plots  was  subsampled  on  the  ground. 
At  the  LS  layer  the  grid  spacing  was  5  km,  at  the  HAP  layer  10  km,  20  km  at  the 

At  each  sample  location  land 
cover  class  was  evaluated. 
Land  cover  polygons  were 
mapped  for  area  estimation  at 
remotely  sensed  locations  and 
were  point  sampled  at  ground 
locations.  Population  totals 
were  estimated  using  several 
estimators,  including  regression 
estimation.  Regression  results 
were  disappointing  in  that 
correlations  between  covariates 
and  independent  variables  were 
not  satisfactory.  A  number  of 
factors,  including  unrecognized 
sources  of  variation  and 
changing  objectives  were 
deemed  responsible.  A 
possible  source  of  uncontrolled 
variation  was  plot  collocation. 
For  regression  estimators  in 
particular,  plots  at  various 
sample  stages  must  be 
accurately  collocated.  This 
study  was  designed  to  investigate  the  hypothesis  that  apparently  minor  plot 
collocation  errors  can  significantly  contribute  to  area  estimation  error. 

Two  factors  that  could  contribute  to  relative  error  magnitude  were  considered;  the 
amount  of  a  given  land  cover  class  on  the  plot  and  the  complexity  of  the  land 
cover  pattern  on  the  plot.  If  a  plot  is  homogenous  with  respect  to  land  cover  class, 
then  small  errors  in  collocation  would,  at  best,  result  in  small  relative  errors  in 
area  estimation.  On  the  other  hand,  if  the  plot  has  small  amounts  of  a  given  land 
cover  class,  then  small  collocation  errors  could  result  in  larger  relative  area 
estimation  errors.  To  represent  land  cover  class  complexity,  the  fractal  dimension 
of  the  mapped  land  cover  pattern  within  the  plot  was  estimated.  The  hypothesis 
then,  was  that  area  estimation  error  for  a  given  land  cover  class  (expressed  as  the 
difference  in  percent  of  plot  area  in  that  land  cover  class  on  a  control  plot  versus  a 


LAP  layer,  and  40  km  on  the  ground,  Figure  1 . 


Figure  1 --Three-phase  sampling  with  subsampling 
grid  design  used  for  multi-resource  inventories  in 
Alaska  during  the  1980's. 


554 


mislocated  plot)  would  be  a  function  of  the  total  plot  area  in  that  land  cover  class 
(expressed  as  a  percent  of  total  plot  area)  and  of  the  overall  fractal  dimension  of 
the  land  cover  pattern  on  the  plot. 

METHODS 
Study  Area 

Field  data  for  this  study  did  not  exist;  this  study  is  a  simulation  in  so  far  as 
production  of  collocation  errors  is  concerned.  Aerial  photos  from  inventories  of 
the  central  interior,  east-central  interior,  and  south  central  coastal  regions  of 
Alaska  (Figure  2.)  were  used  as  the  data  source. 

Data  Collection 


Color  infra-red  aerial  photos 
of  3 1  different  locations  were 
subjectively  selected  to 
represent  a  variety  of 
topographies,  latitudes, 
longitudes,  and  land  cover 
classes.  Aerial  photo  scales 
ranged  from  1:3,000  to 
1:6,000. 

Figure  2--Study  plots  were  selected  from  within 
the  shaded  area  above.  An  aerial  photo  interpreter 

classified  and  delineated,  on 
transparent  overlays,  up  to  4  land  cover  classes  on  each  photo.  The  entire  photo 
was  classified.  Classes  were  chosen  to  simplify  the  classification  process  and 
thereby  reduce  interpretation  error.    The  classes  were  productive  forestland 

Oil 

(forestland  capable  of  producing  1.4  m  .ha"  «yr"  ),  other  forestland,  nonforest,  and 
water. 

Each  of  the  resulting  3 1  land  cover  class  map  overlays  was  digitized.  On  the 
digitized  images  8  ha  control  plots  were  drawn.  These  plots  served  as  the  source 
of  the  true,  or  ground,  area.  For  each  control  plot,  area  by  land  cover  class  was 
measured  and  converted  to  a  relative  value  (Pc). 

Improper  collocation  between  these  control  (or  ground)  plots  and  their 
associated  aerial  photo  plots  was  modeled  by  overlaying  new  plots  that  were  off- 
center.  For  each  control  plot,  improperly  collocated  simulated  aerial  photo  plots 
were  created  by  overlaying,  off-center,  new  plots  moved  in  one  of  eight  cardinal 
directions.  Directions  of  movement  were  randomly  chosen  without  replacement. 


555 


The  amount  of  collocation  error  was  approximately  0.65  cm  on  1:6,000  scale 
photos  (approximately  40  m  on  the  ground).  The  magnitude  of  this  error  derives 
from  field  experience  gained  during  the  course  of  the  inventories.  Five  such 
mislocated  plots  were  established  for  each  control  plot.  For  each  mislocated  plot, 
area  by  land  cover  class  was  also  measured  and  converted  to  a  relative  value  (Pm). 

Sets  of  control  and  moved  plots  were  created  until  there  were  at  least  5 
observations  in  each  cell  of  frequency  tables  for  each  land  cover  class.  The 
frequency  tables  listed  number  of  plots  by  complexity  class  (simple  vs  complex) 
and  percentage  of  the  control  plot  in  the  land  cover  class.  The  percentages  of  the 
control  plot  in  a  given  land  cover  class  were  grouped  into  quarters.  In  order  to 
place  each  plot  in  a  complexity  class,  estimates  of  the  fractal  dimension  of  the 
land  cover  pattern  on  control  plots  were  made  using  the  dividers  method 
(Sugihara  and  May  1990).  Complexity  class  was  established  according  to  a 
vegetation  complexity  model  for  Alaska  (van  Hees  1994).  Ultimately,  195  sets  of 
control  plots  and  associated  moved  plots  were  used  for  this  study. 

Data  Analysis 

For  each  set  of  control  and  moved  plots,  the  differences  in  percent  of  area  (or 
area  estimation  error),  by  land  cover  class,  between  the  moved  and  control  plots 
were  found.  The  mean  of  the  absolute  values  of  the  five  differences  in  percent  of 
area  (Pe)  was  then  calculated  as: 

Pe  =  (X,  IPC  -  PJ)/ 5  /  =  1  to  5.  (1) 

where:  Pc  and  Pm  as  defined  above. 

For  preliminary  analysis,  the  GLM  (General  Linear  Models)  procedure  (SAS 
1990a)  was  used  to  fit  a  linear  regression  model,  by  land  cover  class,  to  the  data. 
The  model  used  was: 

Pe  =   p0  +  Pi  (  Pc  )  +  p2  (  Fd  )  +  p3  (  Pc  *  Fd  )  (2) 

where:  Pe  =  mean  absolute  area  difference, 

po  =  intercept  parameter, 
P  1,2,3  =  regression  coefficients, 
Fd  =  fractal  dimension.       1  <  Fd  <  2 

To  investigate  the  possibility  that  directional  bias  existed,  the  same  general 
linear  model  as  in  (2)  was  used  except  that  individual  absolute  difference  for  the 


556 


particular  direction  of  mislocation  was  the  dependent  variable  rather  than  mean 
absolute  difference  over  the  five  mislocated  plots. 

Subsequent  to  linear  analyses,  cluster  analyses  were  conducted  using  the 
FASTCLUS  procedure  (SAS  1990b).  Using  FASTCLUS,  observations  are 
divided  into  clusters  on  the  basis  of  Euclidean  distances  computed  from  one  or 
more  quantitative  variables.  For  this  study  the  variables  used  for  clustering  were 
Pe,  Pc,  Fd,  and  (Pc  *  Fd). 

RESULTS 

Initial  screening  by  linear  regression  showed  significant  coefficients  for  the 
productive  forestland  class  only.  7-values  and  analysis  of  variance  statistics  are 
presented  in  Tables  1  and  2.    The  predictive  value  of  the  regression  relation 

2  2 

however,  as  measured  by  r  ,  was  negligible.  The  r  value  for  this  relationship  was 
0.129. 


Table  1—r-tests  of  significance  of  individual  regression  coefficients  for  multiple 
linear  regression  analysis  of  Pe  against  Pc,  fractal  dimension  (Fd),  and  Pc*  Fd. 


Variable 

Coefficient 

Standard  error 

T 

P  (two-tail) 

Intercept 

Pc 

Fd 

Pc*Fd 

0.2244931 
-0.3542443 
-0.1525531 

0.3058221 

0.0579952 
0.1019977 
0.0481601 
0.0886174 

3.87 
-3.47 
-3.17 

3.45 

0.0002 
0.0008 
0.0021 
0.0009 

Table  2—Analysis  of  variance  for  multiple  linear  regression  analysis  of  Pe 
against  Pc,  fractal  dimension  (Fd),  and  Pc*  Fd. 

Source  DF 

Sum  of  squares 

Mean  square 

F  value 

Pr>F 

Model  3 
Error  84 
Total  87 

0.00942545 
0.06352763 
0.07295308 

0.00314182 
0.00075628 

4.15 

0.0085 

Adding  a  squared  interaction  term  so  that  the  model  became 

Pe  =   Po  +  Pi  (  Pc  )  +  P2  (  Fd  )  +  p3  (  Pc  *  Fd  )  +  p4  (  Pc  *  Fd  )\  (3) 

improved  predictive  power.  The  r  value  for  the  expanded  model  was  0.375. 
Again,  the  productive  forestland  class  was  the  only  class  for  which  all  regression 
coefficients  were  significant. 

The  mean  difference  (Pe)  between  the  percentages  on  the  control  plots  versus  the 
mislocated  plots  ranged  from  zero  to  10.9  percent  with  a  mean  of  4.3  percent  and 


557 


standard  deviation  of  2.9  percent.  In  the  case  where  there  was  zero  difference 
between  the  control  and  the  mislocated  plots  all  plots  were  homogenous.  Land 
class  pattern  fractal  dimension  (Fd)  ranged  from  1.000  for  homogenous  plots  to 
1.530  for  the  most  complex  patterns.  Mean  Fd  was  1.165  with  a  standard 
deviation  of  0.109. 

Model  (3)  was  used  to  examine  the  data  for  possible  directional  bias.  That  is, 
could  the  direction  in  which  mislocation  of  the  aerial  photo  plot  occurred  affect 
the  magnitude  of  relative  area  estimation  error.  Table  3  presents  the  results  of  this 
analysis.  Noticeable  is  the  strength  of  the  estimated  regressions  for  those 
observations  where  the  direction  of  mislocation  was  along  the  north/northeast  by 
south/southwest  axis  (highlighted  figures  in  Table  3)  when  compared  with  the 
estimates  for  other  directions. 


Table  3~F-values  and  T  -scores  for  estimated  coefficients  of  model  (3)  by  direction  of 
movement  of  improperly  collocated  plots  for  productive  forestland. 


Direction 

df 

F 

T  for  H0  =  0 

r2 

Pc 

Fd 

Pc*Fd 

(  Pc  *  Fd  )2 

North 

4,54 

5.82 

-2.99 

-2.70 

3.70 

-4.37 

.3011 

Northeast 

4,42 

4.02 

-2.59 

-2.61 

3.30 

-1.83 

.2767 

East 

4,57 

2.93 

-1.92 

-1.32 

2.53 

-3.29 

.1706 

Southeast 

4,38 

1.53 

-1.81 

-1.21 

2.28 

-1.68 

.1388 

South 

4,60 

6.12 

-3.31 

-2.53 

4.04 

-3.96 

.2898 

Southwest 

4,30 

4.09 

-1.74 

-1.80 

2.83 

-2.87 

.3526 

West 

4,64 

1.44 

-1.77 

-1.50 

2.01 

-2.08 

.0824 

Northwest 

4,35 

1.38 

-1.12 

-0.77 

1.08 

-1.16 

.1359 

Clustering  of  the  data  was  undertaken  to  examine  the  possibility  that  the 
predictive  power  of  the  model  (using  equation  (1))  could  be  improved. 
FASTCLUS  was  run  setting  the  number  of  clusters  at  3  and  4.  Overall  r2  for  the 
4-cluster  grouping  was  higher  than  for  the  3-cluster  grouping  (0.886  vs  0.829). 
Clusters  were  separated,  by  FASTCLUS,  into  the  following  ranges  of  Pc  with 
almost  no  overlap.  The  groups  of  Pc  were  0  -  15%  (n=18),  15%  -  45%  (n=20), 
45%  -  75%  (n=31),  and  75%  -  100%  (n=19).  Model  (2)  was  then  used  to  examine 
the  within  cluster  regression  relationships.  In  the  first  and  last  clusters  the 
regression  relationships  were  quite  strong  (r  =  0.729  and  0.841  respectively). 
The  relationship  for  the  second  cluster  was  much  weaker  (r2  =  0.385)  and  for  the 
third  cluster  the  regression  relationship  was  essentially  nonexistent  (r  =  0.026). 

DISCUSSION 

With  regards  to  the  productive  forestland  component  of  the  study  area,  the 
hypothesis  that  area  estimation  error  would  be  a  function  of  the  percentage  of  the 


558 


control  plot  (Pc)  in  a  land  cover  class  and  the  fractal  dimension  of  the  land  cover 
pattern  is  supported  by  the  results  of  this  study. 

Although  this  study  does  not  provide  substantive  clues  as  to  why  the  model  was 
more  informative  for  productive  forestland  than  for  other  land  cover  classes  a 
possible  explanation  lies  in  the  physiography  of  the  study  area  and  the  discovery 
of  possible  directional  bias.  Patterns  of  productive  forestland  tend  to  be  linear  in 
the  study  area.  In  interior  Alaska  productive  forestland  is  found  along  rivers  on 
south  facing  slopes  whereas  in  coastal  regions  it  is  "marginal"  -  that  is  along  the 
margins  next  to  water  and  up  to  about  600  m  above  sea  level.  Rivers  and  water 
bodies  in  the  study  area  have  a  predominantly  east/west  orientation.  Thus,  plot 
mislocations  to  the  north  or  south  would  likely  have  more  impact  on  error 
magnitude  than  mislocations  to  the  east  or  west. 

Clustering  of  the  observations  did  improve  predictive  ability  for  certain  ranges 
of  the  data;  particularly  the  upper  and  lower  ends  of  the  control  plot  percentage 
scale.  It  would  be  possible  then,  to  more  accurately  assess  impacts  of  mislocation 
knowing  the  percentage  of  the  plot  in  a  land  cover  class  along  with  an  estimate  of 
fractal  dimension  of  the  mapped  land  cover  classification. 

Although  the  results  of  this  study  do  not  provide  strong  modeling  capabilities 
they  do  provide  useable  guidelines.  It  is  apparent  that  small  mislocations  between 
aerial  photo  and  ground  plots  can  produce  significant  area  estimation  errors  (up  to 
10  percent)  and  that  inventory  designers  and  data  analysts  must  be  aware  of 
physiography  in  order  to  consider  possible  introduction  of  directional  bias.  Also, 
there  may  be  some  predictive  capability  for  extremes  of  land  cover  percentages. 

LITERATURE  CITED 

SAS.  1990a.  SAS/STAT  User's  guide,  Version  6,  Fourth  ed.,  vol.  2.  SAS 

Institute,  Inc.,  Cary,  NC.  USA. 
SAS.  1990b.  SAS/STAT  User's  guide,  Version  6,  Fourth  ed.,  vol.  1.  SAS 

Institute,  Inc.,  Cary,  NC.  USA. 
Schreuder,  Hans  T.,  T.  G.  Gregoire,  and  G.  B.  Wood.   1992.  Sampling  methods 

for  multi-resource  forest  inventory.  John  Wiley  and  Sons,  Inc.,  605  Third  Ave., 

New  York,  New  York.  10158-0012.  446  p. 
Sugihara,  George  &  May,  Robert  M.  1990.  Applications  of  fractals  in  ecology. 

Trends  in  Ecology  and  Evolution,  Elsevier  Science  Publishers  Ltd.,  (UK),  5(3): 

79-86. 

van  Hees,  Willem  W.  S.  1994.  A  fractal  model  of  vegetation  complexity  in 
Alaska.  Landscape  Ecology.  9(4):271-278. 


559 


Sensitivity  of  Landscape  Pattern 
Metrics  to  Misclassification  and 
Differences  in  Land  Cover 
Composition1 

James  D.  Wickham2,  Robert  V.  O'Neill3,  Kurt  H.  Riitters2, 
Timothy  G.  Wade4,  and  K.  Bruce  Jones5 


Abstract.  Calculation  of  landscape  metrics  from  land  cover  data  is 
becoming  increasingly  common.  Some  studies  have  shown  that  these 
measurements  are  sensitive  to  differences  in  land  cover  composition,  but 
none  are  known  to  have  tested  also  their  sensitivity  to  land  cover 
misclassification..  An  error  simulation  model  was  written  to  test  the 
sensitivity  of  selected  landscape  pattern  metrics  to  misclassification,  and 
regression  analysis  was  used  to  determine  if  these  metrics  were 
significantly  related  to  differences  in  land  cover  composition. 
Comparison  of  sensitivity  and  regression  results  suggests  that  differences 
in  land  cover  composition  need  to  be  about  5  percent  greater  than  the 
misclassification  rate  to  be  confident  that  differences  in  landscape  metrics 
are  not  due  to  misclassification. 


1  Forthcoming  in  Photogrammetric  Engineering  &  Remote  Sensing,  1996,  Volume  62. 

2  Tennessee  Valley  Authority,  Norris,  TN  37828. 

3  Oak  Ridge  National  Laboratory,  Oak  Ridge,  TN  37831. 

4  Desert  Research  Institute,  Reno,  NV  89512. 

5  U.S.  Environmental  Protection  Agency,  Las  Vegas,  NV  89119. 


560 


Spatial  (In)consistency  of  Watershed 
Delineations  Among  Agencies  and 
Scales  in  Pennsylvania 

Wayne  L.  Myers1,  Barry  M.  Evans2  &  Michael  C.  Anderson3 


Abstract.--The  Office  for  Remote  Sensing  of  Earth  Resources  at 
Penn  State  University  serves  as  a  GIS  resource  center  and  repository 
for  the  Pennsylvania  Department  of  Environmental  Protection.  One 
of  the  tasks  in  this  context  involves  harmonizing  different  map 
perspectives  on  watersheds  between  agencies  and  scales,  such  as 
USGS  and  the  State  Water  Plan.  This  provides  an  interesting 
opportunity  to  study  the  nature,  degree,  and  distribution  of  spatial 
disparity  in  mappings  of  supposedly  well-defined  and  naturally 
hierarchical  environmental  elements.  Spatial  and  statistical  patterns 
of  discrepancy  in  divides  and  pour  points  are  considered  relative  to 
scale  and  environmental  resource  management  focus. 


INTRODUCTION 

The  Office  for  Remote  Sensing  of  Earth  Resources  (ORSER)  in  the 
Environmental  Resources  Research  Institute  (ERRI)  at  The  Pennsylvania 
State  University  has  a  considerable  history  of  spatial  approaches  to  watershed 
studies  and  hydrologic  modeling.  ORSER  also  serves  as  a  GIS  resource 
center  and  spatial  data  repository  for  the  Pennsylvania  State  Department  of 
Environmental  Protection  (DEP).  The  latter  agency  is  regularly  concerned 
with  water  resource  issues  ranging  in  scope  from  local  up  to  multi-state 
concerns  regarding  Chesapeake  Bay.  These  multifarious  water  resource 
purviews  involve  several  mappings  of  watersheds  at  different  levels  of 
discrimination,  originating  from  different  authorities.  Although  one  usually 
thinks  of  watersheds  as  being  physiographically  determined,  some  degree  of 
inconsistency  among  the  suite  of  mappings  was  evident.   In  the  interest  of 


'Assoc.  Prof,  of  Forest  Biometrics,  Penn  State  Univ.,  Univ.  Park,  PA. 

Senior  Research  Assistant,  Environ.  Resources  Research  Inst,  Penn  State  Univ.,  Univ.  Park,  PA. 
Research  Technician,  Environ.  Resources  Research  InsL,  Penn  State  Univ.,  Univ.  Park,  PA. 


561 


clarity  and  consistency,  ORSER  was  tasked  by  DEP  to  harmonize  the  several 
mappings  in  a  GIS  framework. 

Comparative  analysis  of  habitat  variables  among  and  within  watersheds  is 
one  of  the  undertakings  for  Pennsylvania's  ongoing  Gap  Analysis  of 
conservation  status  for  vertebrates.  A  Gap  successor  project  will  use  digital 
elevation  models  and  stream  reach  files  to  partition  prior  watershed  mappings 
into  "reachsheds"  which  will  then  be  reaggregated  according  to  similarity 
criteria  as  a  basis  for  critical  area  assessment  of  landscapes.  Because  of  these 
research  dependencies  on  accuracy  and  consistency  of  watershed  files,  it  was 
decided  to  conduct  some  accuracy  investigations  in  conjunction  with  the 
harmonization  effort. 

PROCEDURE 

Three  different  watershed  data  series  were  selected  for  the  present 
comparative  analysis.  USGS  digital  drainage  basin  boundaries  of  named 
streams  in  Pennsylvania  (Hoffman  &  Bubb,  1995)  comprised  the  lower  tier  of 
(smallest)  watersheds.  Watersheds  recognized  by  the  State  Water  Plan  of  the 
Pennsylvania  Bureau  of  Water  Resources  Management  (1:250,000  scale) 
constituted  the  middle  tier.  There  are  a  total  of  104  such  watersheds 
delineated  for  the  state.  USGS  8-digit  hydrologic  unit  boundaries  (1:250,000 
scale)  represented  the  top  tier  of  (largest)  watersheds.  Analysis  was  restricted 
to  top-tier  watersheds  lying  entirely  within  the  boundaries  of  the 
Commonwealth.  Outlines  of  these  23  watersheds  are  shown  in  Figure  1. 

The  26  middle-tier  watersheds  comprising  the  Susquehanna  River  Basin 
where  used  in  the  first  stage  of  analysis.  Outlines  of  these  26  watersheds  are 
shown  in  Figure  2.  The  3,585  lower-tier  watersheds  for  the  Susquehanna 
Basin  (not  shown)  were  aggregated  upward  to  match  the  corresponding  26 
middle-tier  watersheds.  Matched  watersheds  were  then  compared  for 
discrepancies  in  boundaries  and  pour  points  using  Arc/Info  software. 

All  of  the  whole  watersheds  for  both  coverages  (smaller  scale  and 
aggregated)  were  extracted  and  placed  in  their  own  coverages.  They  were 
both  numbered  the  same  way,  i.e.,  watershed-id  #1  represents  the  same 
watershed  in  both  coverages.  The  mean  area  and  mean  perimeter  of  each 
pair  matched  by  id#  were  then  computed.  A  union  operation  was  then 
performed  on  the  two  coverages.  The  number  of  slivers  and  sliver  area  were 
determined  for  each  pair.  An  index  of  sliver  width  was  developed  as  the  sliver 
area  divided  by  the  average  perimeter. 


562 


563 


564 


For  automated  determination  of  pour  points,  nodes  were  placed  at  100 
meter  intervals  along  the  boundaries,  and  elevations  transferred  from  a 
corresponding  100-meter  DEM.  The  DEM  was  a  3  arc-second  USGS  DEM 
resampled  on  an  Albers  equal-area  projection  used  for  the  watersheds.  Pour 
points  were  determined  as  the  minimum  elevation  on  the  boundary. 

In  the  second  stage  of  analysis,  middle-tier  watersheds  were  similarly 
aggregated  to  match  the  23  top-tier  watersheds.  Metrics  of  discrepancy  were 
determined  in  the  same  manner  as  for  the  first  stage  of  analysis. 

RESULTS 

Table  1  contains  detailed  results  for  aggregating  from  smallest  watersheds 
to  mid-level  watersheds. 


Table  1.  Results  of  aggregating  to  mid-level  (from  smallest)  watersheds. 


Pour  pt . 

Avr.  perim. 

Avr.  area 

No.  of 

Sliver 

Sliver 

ID 

dist (m) 

(km) 

(ha) 

slivers 

area(ha) 

width  (m) 

1 

36,441 

211.5 

142,920 

95 

7,198 

340 

2 

2,815 

344.7 

253,958 

195 

10,419 

302 

3 

3 , 653 

214  . 0 

122 , 677 

103 

8,002 

374 

4 

165 

186.4 

106,361 

84 

7,928 

425 

5 

431 

231.9 

93,026 

89 

11,040 

476 

6 

558 

285.4 

133,025 

138 

10,362 

363 

7 

344 

317.7 

269,886 

170 

10,901 

343 

8 

561 

279.5 

202,310 

146 

8,013 

287 

9 

3,033 

240.6 

127,934 

134 

8,658 

360 

10 

4,828 

242.7 

134,558 

121 

10,324 

425 

11 

14,329 

255.6 

116,963 

112 

10,624 

416 

12 

887 

182.3 

105,931 

58 

7,752 

425 

13 

604 

187.1 

99,741 

102 

5,039 

269 

14 

1,997 

212.4 

90,732 

45 

12,731 

599 

15 

330 

263.9 

124,220 

119 

9,513 

361 

16 

561 

295.4 

199,776 

103 

9,866 

334 

17 

5,295 

293.6 

161,010 

122 

14,447 

492 

18 

5,124 

243.8 

129,307 

110 

12,562 

515 

19 

1,136 

294.9 

151,060 

96 

10,608 

360 

20 

453 

259.9 

192,572 

108 

9,602 

369 

21 

861 

170.2 

63,936 

62 

5,434 

319 

22 

302 

291.6 

145,584 

134 

6,962 

239 

23 

683 

215.7 

148,275 

99 

6,012 

279 

24 

2,162 

198.3 

77,829 

68 

5,461 

275 

25 

3,613 

265.3 

131,006 

97 

9,079 

342 

26 

11,877 

184.2 

64,804 

67 

9,577 

520 

565 


Table  2  contains  results  of  aggregating  to  highest  level  (largest)  watersheds 
from  mid-level  watersheds. 


Table  2.  Results  of  aggregating  to  highest  level  (largest)  watersheds. 


Pour  pt. 

Avr.  perim. 

Avr .  area 

NO.  Of 

Sliver 

Sliver 

ID 

dist (m) 

(km) 

(ha) 

slivers 

area(ha) 

width  (m) 

1 

191 

423.2 

512,086 

107 

33,934 

802 

2 

2,336 

294.2 

253,699 

83 

11,990 

408 

3 

4,868 

442.9 

436,231 

79 

29,153 

658 

4 

140 

500.2 

463,004 

116 

26,806 

536 

5 

719 

198.6 

151,596 

38 

11,749 

592 

6 

926 

280.7 

271,222 

66 

15,189 

541 

7 

1,942 

344.1 

322,189 

52 

22,876 

665 

8 

1,723 

246.5 

202,984 

59 

11,144 

452 

9 

140 

379.3 

471,123 

101 

15,941 

420 

10 

1,066 

217.4 

218,444 

63 

10,243 

471 

11 

5,537 

366.9 

437,091 

68 

28,285 

771 

12 

2,939 

389.2 

353,314 

92 

19,726 

507 

13 

926 

327.5 

416,670 

68 

15,792 

482 

14 

1,723 

265.1 

199,032 

51 

11,132 

420 

15 

920 

429.3 

375,489 

99 

20,543 

479 

16 

677 

198.3 

127,887 

36 

14,501 

731 

17 

2,079 

530.2 

496,632 

123 

32,370 

611 

18 

628 

271.4 

256,782 

54 

14,137 

521 

19 

920 

400.6 

377,071 

82 

23,534 

587 

20 

7,532 

297.2 

351,141 

43 

24,995 

841 

21 

5,145 

440.5 

478,482 

73 

39,309 

892 

22 

403 

228.5 

133,976 

42 

20,869 

913 

23 

455 

340.0 

246,702 

53 

25,012 

736 

The  area  in  slivers  after  aggregation  reflects  the  uncertainty  of  boundary 
delineation  between  scale  levels  for  the  watersheds.  Pooling  the  sliver  areas 
and  mean  areas  over  all  watersheds,  the  sliver  area  represents  6.6%  of  the 
mean  area.  The  corresponding  percentage  for  aggregating  from  mid-level  to 
large  watershed  level  is  6.3%.  These  percentages  are  surprisingly  consistent, 
and  thus  indicative  that  the  6%  figure  holds  across  scales  for  Pennsylvania 
watershed  demarcation. 

The  observed  distance  between  aggregated  and  original  pour  points  is  quite 
another  matter.  This  distance  differential  is  highly  variable,  ranging  from  less 
than  200  meters  to  more  than  7  km  for  scaling  up  from  mid-level  to  large 
watersheds.  The  variability  is  even  greater  for  scaling  up  small  watersheds  to 
mid-level  watersheds. 


566 


Upon  investigation,  the  36+  km  figure  for  watershed  id#l  in  the  latter  set 
constitutes  an  anomaly.  In  this  particular  case,  the  pour  points  before  and 
after  aggregation  were  placed  at  opposite  ends  of  the  watershed.  This  is 
presumably  attributable  to  inconsistency  of  the  DEM  used  for  pour  point 
location. 

Irregularities  in  the  DEM  undoubtedly  also  contribute  to  large  pour  point 
discrepancies  in  the  remaining  cases,  but  other  factors  are  also  involved.  The 
sliver  width  index  is  one  measure  of  mean  sliver  width.  This  is  on  the  order 
of  one-half  kilometer,  being  somewhat  less  for  the  lower  aggregation  and 
somewhat  more  for  the  upper  aggregation.  This  would  account  for  a 
corresponding  order  of  difference  in  pour  points. 

Constrained  reaches  are  particularly  problematic  relative  to  location  of  pour 
points.  Such  reaches  give  rise  to  "corridor"  watersheds  having  very  limited 
area,  but  causing  substantial  displacements  of  pour  points  depending  on 
whether  or  not  they  are  treated  separately.  There  are  also  more  arbitrary 
instances  such  as  cases  where  delineations  track  parts  of  boundaries  for  civil 
divisions. 

CONCLUSIONS 

Multi-scale  analysis  involving  spatially  explicit  datasets  from  different 
sources  at  different  levels  of  detail  will  require  that  inconsistencies  be 
harmonized  across  scales.  The  process  of  harmonization  affords  opportunity 
for  comparative  analysis  of  inconsistencies  to  determine  relative  orders  of 
spatial  accuracy. 

Comparative  analysis  of  watershed  delineations  at  three  scales  in 
Pennsylvania  reveals  that  differing  positions  for  drainage  divides  induce  errors 
corresponding  to  approximately  6%  of  watershed  area.  In  contrast,  the  range 
of  difference  in  pour  point  locations  spans  more  than  an  order  of  magnitude. 

REFERENCES 

Hoffman,  S.  A.  and  J.  E.  Bubb.  1995.  Digital  drainage  basin  boundaries  of 
named  streams  in  Pennsylvania.  U.S.  Geological  Survey,  Open-File  Report 
95 -xxx,  Lemoyne,  PA. 


567 


COMMUNICATING  THE  RESULTS 
OF    ACCURACY  ASSESSMENT: 
METADATA,  DIGITAL  LIBRARIES, 
AND  ASSESSING  FITNESS  FOR  USE 

Michael  F.  Goodchild1 


The  concept  is  proposed  of  a  data  life  cycle  extending  from  the  initial 
collection  of  data  in  the  field  or  by  remote  sensing  to  eventual  archiving. 
Although  GIS  has  been  perceived  to  date  largely  as  an  analytic 
technology  confined  to  a  small  part  of  the  data  life  cycle,  its  influence  is 
increasingly  seen  as  extending  throughout  the  cycle,  in  the  form  of  newer 
technologies  such  as  field  GIS  and  digital  spatial  data  libraries.  The 
concept  of  a  data  model  is  introduced  and  examples  are  given  of  the  role 
of  data  modeling  along  the  life  cycle.  Accuracy  assessment  and  concern 
for  broader  issues  of  data  quality  are  critical  if  data  are  to  be  assessed  as 
to  fitness  for  use  by  the  various  actors  in  the  life  cycle.  GIS  data  models 
must  be  extended  to  accommodate  essential  information  on  accuracy. 
The  role  of  accuracy  assessment  in  metadata  is  discussed,  and  a  new 
view  of  metadata  is  proposed  with  extended  functions  and  a  hierarchical 
structure. 


Professor  of  Geography  and  Director,  National  Center  for  Geographic  Information  and 
Analysis,  University  of  California,  Santa  Barbara,  CA  93106-4060.  Email: 
good®ncgia . ucsb . edu 


568 


Covariate-Directed  Sampling  for 
Assessing  Species  Richness 


G.P.  Patil,  Glen  Johnson 
and  Matteo  Grigoletto1 


Abstract. — Since  species  richness  is  a  spatially  non-additive  vari- 
able, it  can  not  be  estimated  with  conventional  moment  estimators. 
We  may,  however,  exploit  the  species-area  relationship  which  im- 
plies that  the  number  of  species  increases  according  to  a  law  of  di- 
minishing returns  as  the  sampled  area  increases.  An  efficient  sam- 
pling method  would  then  maximize  the  number  of  species  encoun- 
tered within  a  fixed  amount  of  area  that  can  be  affordably  sam- 
pled. We  suggest  that  when  spatial  covariate  information  is  avail- 
able, it  should  be  exploited  for  directing  the  location  of  sample 
units  in  such  a  manner  that  increases  the  habitat  diversity  observed 
within  a  minimum  number  of  sample  units.  This  approach  was  eval- 
uated through  a  retrospective  assessment  of  breeding  bird  species 
richness  in  Pennsylvania  using  cumulative  tree  richness  as  the  co- 
variate. Working  with  EMAP  hexagons  (635  square  kilometers)  as 
the  primary  sampling  unit,  we  found  that  although  tree  richness 
was  a  fairly  weak  covariate,  it  still  outperformed  random  sampling. 


INTRODUCTION 

With  such  a  strong  and  legitimate  concern  for  the  alarming  loss  of  biodi- 
versity on  our  planet  (Wilson,  1988;  Stevens,  1995),  monitoring  methods  are 
essential  for  quantifying  the  biodiversity  of  large  geographic  regions.  With  all 
the  shortcomings  of  "indices",  actual  species  richness  (the  number  of  different 
species)  appears  to  be  the  least  controversial  measure  of  biodiversity.  Mean- 
while, biodiversity  researchers  recognize  that  reliable  methods  of  estimation 
still  require  development  (Yoon,  1995). 

Since  species  richness  is  a  spatially  non-additive  variable,  we  can  not  esti- 
mate its  total  from  conventional  moment  estimators,  such  as  by  multiplying  a 
sample  mean  by  the  number  of  population  units  in  the  sampling  frame.  For 
this  reason,  we  turn  to  species-area  curves,  which  are  used  by  ecologists  for 
several  reasons  (Kilburn,  1966),  including  the  prediction  of  species  richness  in 
larger  areas  than  those  sampled  (Evans,  Clark  and  Brand,  1955). 

Center  for  Statistical  Ecology  and  Environmental  Statistics,  Department  of  Statistics, 
Penn  State  University,  University  Park,  PA  16801 


569 


The  species-area  relationship  basically  states  that  as  the  area  within  a  ho- 
mogeneous habitat  increases,  the  number  of  different  species  encountered  will 
also  increase  until  "a  point  of  no  return" ,  after  which  increasing  the  area  does 
not  further  increase  the  number  of  different  species  encountered.  The  com- 
mon model  for  this  process,  as  originally  proposed  by  Arrhenius  (1921),  is  a 
power  function,  presented  as  S  —  kAz,  where  S  is  the  species  richness  and  A 
is  the  area,  while  z  and  k  are  population  specific  parameters.  For  applications 
to  wildlife  conservation  issues,  see  Usher  (1985).  Although  the  power  func- 
tion has  been  traditionally  used  to  model  the  species-area  relationship  within 
homogeneous  habitat,  Johnson  and  Patil  (1995)  observed  a  classical  power 
function  response  for  breeding  bird  richness  across  the  whole  state  of  Pennsyl- 
vania which  encompasses  very  heterogeneous  habitat.  If  the  power  function  is 
fit  from  sample  data,  its  ability  to  extrapolate  to  a  larger  area  is  limited  by 
upward  bias  since  this  model  is  unbounded  (Williams,  1995). 

Our  objective  is  to  develop  a  sampling  plan  that  maximizes  the  acceleration 
of  a  species-area  curve  towards  it's  plateau  in  order  to  encounter  the  most 
species  within  a  sampled  area.  This  is  especially  critical  when  sampling  from 
a  very  large  geographic  area  like  the  state  of  Pennsylvania  since  this  can  rapidly 
become  a  very  expensive  exercise. 

SAMPLING  STRATEGIES 

When  constructing  a  species-area  curve  from  successive  aggregation  of  dis- 
crete sample  units,  the  usual  approaches  are  to  either  combine  sample  units 
in  a  continuous  fashion  or  to  combine  units  that  are  obtained  at  random  from 
throughout  the  region  of  interest. 

When  habitat  is  diverse  across  the  region,  random  aggregation  may  result 
in  a  steeper  curve  than  is  obtained  from  continuous  aggregation  because  spa- 
tially discontinuous  sample  units  may  encounter  more  diverse  habitats,  there- 
fore increasing  the  chance  of  encountering  different  species.  The  expected 
number  of  species  encountered  in  n  sample  units  of  equal  size  obtained  at 
random,  E[Sn],  can  be  readily  computed  (Kobayashi,  1979)  for  providing  a 
benchmark  to  compare  other  sampling  protocols. 

An  alternative  approach  to  continuous  aggregation  or  random  sampling  is 
to  perform  directed  sampling  based  on  values  of  some  covariate  that  are  readily 
available  for  the  sample  units.  With  the  advent  of  geographic  information  sys- 
tems,we  feel  that  such  covariate  information  is  becoming  more  readily  available 
for  geographic  areas  at  the  landscape  scale  and  above.  The  desired  property 
of  a  covariate  would  be  to  direct  sampling  in  a  manner  that  accelerates  the 
species-area  curve  faster  than  would  be  observed  with  spatially  continuous  or 
random  sampling. 

BREEDING  BIRDS  IN  PENNSYLVANIA 


570 


65 

lo 

94 

to 

Number  of  Species 

73  to  82       |!j  91  to  100    ^  10«  to  119  H  127  lo  139  Q  UMng 

100  to   109  »&|  11S  to   127  H  139  to  14! 


Figure  1:  Bird  richness  in  the  hexagons. 

We  evaluated  the  proposed  approach  of  covariate-directed  sampling  through 
a  retrospective  study  of  a  known  community  of  Breeding  birds  in  Pennsylva- 
nia. Our  database  is  described  in  Johnson  and  Patil  (1995).  The  sampling 
frame  consists  of  a  tessellation  of  Pennsylvania  by  hexagons,  each  635  km2, 
of  the  Environmental  Protection  Agency's  Environmental  Monitoring  and  As- 
sessment Program  (EMAP).  Associated  with  each  hexagon  are  species  lists 
for  breeding  birds,  other  vertebrate  groups  and  trees.  While  other  groups  are 
based  on  records  of  occurrence,  the  breeding  birds  are  based  on  the  much  more 
thorough  Pennsylvania  Breeding  Bird  Survey  (Brauning,  1992).  The  distribu- 
tion of  species  richness  for  breeding  birds  with  respect  to  EMAP  hexagons  is 
displayed  in  Figure  1  in  the  form  of  a  greyscale  thematic  map. 

Of  the  information  available  in  our  database,  tree  species  presented  the 
most  promising  covariate  for  choosing  an  optimal  hexagon  ordering  for  ulti- 
mately measuring  bird  species  richness.  We  basically  hypothesized  that  dif- 
ferences in  bird  species  are  likely  to  be  associated  with  differences  in  tree 
species;  therefore,  if  hexagons  are  chosen  in  an  order  that  corresponds  to  max- 
imum acceleration  of  the  tree  species  richness  curve,  using  this  same  ordering 
will  accelerate  the  bird  richness  curve.  Constructing  the  optimal  tree  species 
richness  curve  was  performed  by  choosing  the  first  hexagon  as  the  one  con- 
taining the  highest  tree  species  richness.  After  noting  which  species  were  in 
the  first  hexagon,  all  members  of  these  species  were  deleted  from  the  remain- 
ing hexagons.  The  second  hexagon  was  then  chosen  as  the  one  containing 
the  highest  tree  richness.  Steps  1  and  2  were  then  repeated  until  all  the  tree 


571 


•      first  directed  cycle 

V     mean  for  random  sampling  after  first  cycle 
o      mean  for  random  sampling 

0  50  100  150 

Hexagons  Observed 

Figure  2:  Species-area  curve  from  tree-directed  sampling  for  one  cycle, 
followed  by  the  expected  value  from  subsequent  random  sampling.  The 
expected  curve  from  completely  random  sampling  is  also  shown. 

species  had  been  accounted  for. 

We  discovered  that  all  tree  species  were  accounted  for  within  the  first  nine 
hexagons  sampled.  At  this  point  we  experimented  with  two  techniques.  The 
first  one  is  to  randomly  sample  part  of  the  remaining  hexagons.  The  second 
technique  is  to  reintroduce  all  of  the  tree  species  that  are  within  the  unsampled 
hexagons,  repeating  the  procedure  used  for  the  first  cycle,  and  so  on  for  a 
certain  number  of  cycles  (this  is  called  a  completely  directed  procedure).  At 
each  step  the  new  bird  species  found  in  the  hexagon  were  recorded.  The 
completely  directed  procedure  proved  to  perform  somewhat  better,  as  seen  in 
Figures  2  and  3  which  provide  results  over  the  whole  state. 

For  the  protocol  which  sampled  the  first  nine  hexagons  by  tree-directed 
sampling,  we  fit  the  power  function  model,  via  linear  regression,  for  additional 
random  samples  of  20,  30  and  50.  The  results  are  presented  in  Table  1,  where 
extrapolation  to  the  total  area  (statewide)  appears  to  yield  unacceptably  high 
bias.  Since  we  are  using  a  covariate  (number  of  tree  species)  for  the  first  cycle 
followed  by  random  sampling,  then  the  number  of  new  tree  species  considered 
increases  just  for  the  first  nine  hexagons,  and  it  becomes  constant  thereafter 
(just  the  area  increases).  To  take  this  into  account,  we  might  use  a  model  of 
the  form  S  =  kAzTp,  where  T  is  the  number  of  tree  species  and  p  is  a  new 
parameter.  Such  a  model  has  been  used  by  Rafe  et  al.  (1985);  what  they  call 
habitat  heterogeneity  in  our  case  is  represented  by  number  of  tree  species.  The 
extrapolated  curves  for  both  models  are  compared  in  Figure  4  for  the  sampling 
protocol  of  one  tree-directed  cycle,  followed  by  20  randomly  chosen  hexagons. 


572 


Figure  3:  Species-area  curve  from  completely  tree-directed  sampling,  with 
demarcation  of  each  sampling  cycle.  The  expected  curve  from  completely 
random  sampling  is  also  shown  for  comparison. 


Table  1:  Extrapolated  species  richness  (EX)  for  one  tree-directed  cycle 
(9  hexagons)  plus  subsequent  random  samples  of  size  20,  30  and  50.  Bias 
equals  the  extrapolated  minus  the  total  known  species. 


Sample  size 

 X  1 

k 

z 

EX 

bias 

29 

36.87 

0.157 

230 

+52 

39 

42.76 

0.139 

216 

+38 

59 

51.42 

0.117 

202 

+24 

Here  we  see  a  substantial  reduction  in  bias  from  incorporating  the  covariate. 

The  main  drawback  of  a  model  which  incorporates  the  covariate  is  that 
there  is  no  trivial  way  to  use  it  when  covariate-directed  selection  of  all  sample 
units  comes  into  play  (completely  directed  procedure).  There  are  no  clear 
values  of  the  variable  T  to  be  used  in  the  second  and  the  following  cycles. 
This  problem  is  important  since  we  obtain  the  steepest  species-area  curve 
with  the  completely  directed  procedure. 

Besides  tree  species,  we  also  had  available  information  on  (i)  fish,  (ii)  mam- 
mals, (iii)  reptiles  and  amphibians  and  (iv)  butterflies  and  skippers.  Using 
these  covariates  we  reapplied  the  same  analysis  as  with  trees,  but  did  not 
discover  a  stronger  covariate  than  tree  richness.  Results  from  studying  these 
other  covariates  can  be  found  in  Grigoletto,  et  al.  (1995). 

DISCUSSION 

Most  of  the  cycle  lengths  in  Figure  5  vary  little  around  10  hexagons.  Why 


573 


0  50  100  150 


Hexagons  Observed 

Figure  4:  Extrappolating  curves  for  Model  1  (S  =  kAz)  and 
Model  2  (S  =  kAzTp). 


Figure  5:  Observed  hexagons  and  unencountered  species  after  4  cycles. 


574 


does  that  happen?  If  we  look  at  it  is  clear  that  the  observed  sites  tend  to 
cluster.  Since  we  are  using  tree-  directed  sampling,  this  happens  because  if 
there  are  certain  areas  consisting  of  multiple  hexagons  with  a  high  number  of 
tree  species,  then  in  different  cycles  we  tend  to  re-observe  these  areas.  In  fact, 
the  larger  clusters  do  contain  hexagons  from  each  of  the  four  cycles. 

After  four  cycles,  only  six  bird  species  remain  unencountered.  In  Figure  5, 
we  see  that  most  of  the  unencountered  species  are  rare  and  appear  just  in 
one,  two  or  three  hexagons.  Pine  Siskin  is  an  exception  since  this  bird  is 
quite  spread  over  central  northern  Pennsylvania.  Since  this  region  is  a  mostly 
forested  area,  this  seems  to  suggest  that  if  the  "area  covered  by  forest"  could 
be  used  jointly  with  tree  richness,  we  would  have  a  stronger  covariate. 

The  primary  purpose  of  this  paper  is  to  suggest  the  idea  of  covariate- 
directed  sampling  for  achieving  observational  economy  when  sampling  to  es- 
timate species  richness,  a  spatially  non-additive  variable.  We  illustrated  the 
approach  with  a  retrospective  study  of  breeding  birds  in  Pennsylvania,  where 
the  database  at  our  disposal  provided  species  lists  for  several  other  commu- 
nities. Of  these,  tree  richness  appeared  to  be  the  best  available  covariate. 
Although  tree  richness  was  not  strongly  correlated  with  bird  species  richness, 
tree-directed  sampling  did  accelerate  the  bird  species-area  curve  faster  than 
was  expected  with  completely  random  sampling.  Therefore,  if  other  covariates 
can  be  obtained  that  are  more  strongly  correlated  with  bird  richness,  then  such 
covariates  are  expected  to  further  improve  the  sampling  efficiency. 

An  outstanding  question  that  still  remains  is  how  to  estimate  species  rich- 
ness statistically,  in  a  manner  that  allows  construction  of  confidence  bounds. 
Progress  has  been  made  (Bunge  and  Fitzpatrick,  1993)  for  the  case  when  pop- 
ulation densities  are  also  measured  within  each  sampled  species;  however,  we 
were  faced  with  presence/absence  responses  for  each  species,  where  the  data- 
analytic  approach  appears  to  be  the  only  alternative. 

Acknowledgments 

Prepared  with  partial  support  from  the  United  States  Environmental  Pro- 
tection Agency,  Environmental  Monitoring  and  Assessment  Program,  EMAP 
Design  and  Statistics  Group  under  a  Cooperative  Agreement  Number  CR- 
821783.  The  contents  have  not  been  subjected  to  Agency  review  and  therefore 
do  not  necessarily  reflect  the  views  of  the  Agency  and  no  official  endorsement 
should  be  inferred. 

REFERENCES 

Arrhenius,  0.  1921.  Species  and  area.  Journal  of  Ecology,  9:95-99. 

Brauning,  D.W.  1992.  Atlas  of  Breeding  Birds  in  Pennsylvania.  University 
of  Pittsburgh  Press,  Pittsburgh.  484  pp. 


575 


Bunge,  J.  and  Fitzpatrick,  M.  1993.  Estimating  the  number  of  species:  a 
review.  Journal  of  the  American  Statistical  Assoc.,  88:364-373. 

Evans,  F.C.,  Clark,  P.J.  and  Brand,  R.H.  1955.  Estimation  of  the  number  of 
species  present  on  a  given  area.  Ecology,  36:342-343. 

Grigoletto,  M.,  Johnson,  G.,  Patil,  G.P.  and  Taillie,  C.  1995.  Using  Covariate- 
Directed  Sampling  of  EMAP  Hexagons  to  Assess  the  Statewide  Species 
Richness  of  Breeding  Birds  in  Pennsylvania.  Technical  Report  no.  95- 
1102.  Center  for  Statistical  Ecology  and  Environmental  Statistics,  De- 
partment of  Statistics,  Penn  State  University,  University  Park,  PA. 

Johnson,  G.D.  and  Patil,  G.P.  1995.  Estimating  statewide  species  richness 
of  breeding  birds  in  Pennsylvania.  Coenosis,  10(2-3):81-87. 

Kilburn,  P.D.  1966.  Analysis  of  the  species-area  relation.  Ecology,  47:831- 
843. 

Kobayashi,  S.  1979.  Species-area  curves,  in  Ord,  J.K.,  Patil,  G.P.  and  Tail- 
lie,  C  (eds.),  Statistical  Distributions  in  Ecological  Work,  pp.  349-368. 
International  Co-operative  Publishing  House,  Fairland,  Maryland. 

Rafe,  R.W.,  Usher,  M.B  and  Jefferson,  R.G.  1985  Birds  on  reserves:  the 
influence  of  area  and  habitat  on  species  richness.  Journal  of  Applied 
Ecology,  22:327-335. 

Stevens,  W.K.  1995.  How  many  species  are  being  lost?  Scientists  try  new 
yardstick.  The  New  York  Times,  p.  C4,  July  25. 

Usher,  M.B.  1985.  Implications  for  the  species-area  relationship  for  wildlife 
conservation.  Journal  of  Environmental  Mngt.,  21:181-191. 

White,  D.,  Kimerling,  A.J.  and  Overton,  W.S.  1992.  Cartographic  and  geo- 
metric components  of  a  global  sampling  design  for  environmental  moni- 
toring. Cartographic  and  Geographic  Information  Systems,  19(1 ) :5— 22. 

Williams,  M.R.  1995.  An  extreme-value  function  model  of  the  species  inci- 
dence and  species-area  relations.  Ecology,  76(8):2607-2616. 

Wilson,  E.O.  1988.  Biodiversity.  National  Academy  Press,  Washington, 
D.C.,  521  pp. 

Yoon,  C.K.  1995.  Monumental  inventory  of  Costa  Rican  forest's  insects  under 
way.  The  New  York  Times,  p.C4,  July  11. 


576 


Data  Accuracy  to  Data  Quality: 

Using  spatial  statistics  to  predict  the  implications  of 
spatial  error  in  point  data 


 Adam  Lewis  and  Michael  F.  Hutchinson 

Abstract.-  Data  error  has  received  a  good  deal  of  attention,  however,  data 

error  alone  is  not  very  informative.  To  judge  the  value  of  a  dataset  for  a 

specific  application,  measures  of  data  quality  are  needed. 

A  quantitative  approach  to  the  problem  of  estimation  of  data  quality  for 

point  data  is  presented,  based  on  the  use  of  geostatistics  to  characterise 

the  GIS  datasets  with  which  the  points  are  to  be  overlaid. 

Results  demonstrate  that  the  intended  application  of  data  is  critical  to  the 

assessment  of  data  quality,  and  that  the  suitability  of  data  for  a  given 

application  can  be  assessed  with  only  a  basic  model  of  the  absolute  spatial 

error  associated  with  the  point  data. 


INTRODUCTION 

A  substantial  body  of  literature  exists  on  the  subject  of  error  in  spatial 
databases. 

For  point  data,  spatial  error  can  be  modelled  as  a  random  process.  Most 
simply  a  bivariate  normal  distribution  can  be  used,  but  other  symmetric  probability 
density  functions  have  also  been  used,  eg.  Bolstad  et  al.  (1990).  The  scaling 
parameter  of  the  distribution,  such  as  the  standard  deviation,  may  be  referred  to  as 
epsilon.  An  alternative  to  the  probabilistic  model  is  the  detenriinistic  interpretation, 
in  which  all  points  are  considered  to  fall  within  a  distance,  epsilon,  of  the  true 
location.  This  extends  to  the  epsilon-band  model  of  spatial  error  in  linear  features. 
These  interpretations  of  spatial  error  are  often  quoted  as  a  standard  in  mapping 
organisations  (Goodchild  1988),  for  instance  "90%  of  points  are  within  25  metres 
of  the  correct  location". 

Attribute  errors  arise  because  thematic  maps,  DEMs  and  other  spatial  datasets 
are  approximations.  The  polygon  representation  of  spatial  variability  maps 
homogeneous  units  with  distinct  boundaries,  however  it  is  well  understood  that  for 
natural  resource  data  these  polygons  are  heterogeneous.  Where  continuous  spatial 
variation  is  represented  on  a  grid  or  lattice,  as  with  a  DEM,  there  is  a  residual  error 
in  the  model,  which  may  be  expressed  as  the  RMS  error. 

Both  spatial  and  attribute  errors  are,  typically,  spatially  autocorrelated;  the 
error  in  one  location  is  dependent  on  the  errors  nearby.  This  reduces  the 
consequences  of  error  in  some  circumstances  -  digital  contours  do  not  cross 
randomly  although  they  may  be  within  each  others  epsilon-band  -  but  complicates 
the  modelling  of  errors.  The  probabilistic  point-epsilon  model  of  spatial  errors 
cannot  be  extended  to  lines  without  complication,  and  models  of  attribute  error 
require  simulation  of  autocorrelated  processes  (Goodchild  et  al.  1992). 


577 


Data  error  versus  data  quality 


Few  studies  have  focussed  on  the  practical  implications  of  errors  in  spatial 
data.  Ultimately  the  question  is  not  the  absolute  error  associated  with  a  dataset, 
but  whether  the  datasets  available  are  of  adequate  quality  to  support  a  given  use. 

This  paper  develops  the  relationship  between  locational  accuracy,  and  data 
quality,  the  ultimate  aim  being  to  provide  a  basis  from  which  to  make  informed 
statements  on  accuracy. 


The  analysis  is  limited  to  spatial  error  of  point  locations  using  a  simple  error 
model.  The  consequence  of  spatial  error  on  the  outcome  of  overlay  with  datasets 
representing  continuous  variables  is  analysed.  The  suitability  of  the  point  data  for 
overlay  with  a  particular  variable  is  assessed  using  the  expected  R2  value  between 
the  outcome  of  overlay  given  the  spatial  error,  and  the  outcome  of  the  overlay  if 
the  points  were  indeed  accurately  placed.  This  provides  a  measure  of  the  quality  of 
the  point  dataset  for  a  particular  use. 

Quadrat  locations  (source  1)  were  recorded  in  a  floristic  database  (Turner  and 
Muek  1992)  in  decimal  degrees,  calculated  from  map  coordinates  read  from 
1:100,000  scale  topographic  maps.  Source  1  coordinate  error  is  assumed  to  be  due 
incorrect  transcription  of  points  onto  maps,  incorrect  reading  of  coordinates  from 
maps,  and  incorrect  transcription  of  coordinate  values. 

Quadrats  locations  (source!)  were  also  marked  onto  1:25,000  topographic 
maps  of  dubious  cartographic  lineage.  It  is  assumed  that  the  quadrat  locations  are 
marked  correctly  onto  the  topographic  maps  relative  to  local  stream,  ridge  and 
road  features  identified  on  the  map.  Source!  coordinate  error  is  assumed  to  be  due 
to  stretches  and  shifts  in  the  low  quality  map  base. 

Source3  data  were  compiled  by  a  separate  process  which  avoids  the  sources 
of  error  associated  with  source  1  and  source!.  These  are  considered  to  be  the  true 
quadrat  locations.  Given  the  cartographic  scale  of  the  maps  and  the  limits  on  the 
accuracy  with  which  a  point  can  be  located  on  a  map  from  a  field  survey,  a 
planimetric  error  of  20  metres  is  regarded  as  very  good. 
Quadrat  locations  are  illustrated  in  figure  1. 


DATA  SOURCES  AND  METHODS 


.  cell  size)  was  interpolated  from 

\  1:25,000  contours  (10  metre 

.  contour  interval)  for  the  study 

•  area       using  ANUDEM 


(Hutchinson  1988,  1989).  Two 
attributes  of  the  terrain  model, 
slope  and  elevation,  were 
estimated  at  each  quadrat 
location  using  the  GIS.  Slope 
was  estimated  using  the  method 
of  Burrough  (1986)  as 
implemented  in  ARCINFO. 


quadrats. 


Figure  1.  Location  of  study 


A  terrain  model  (15  metre 


578 


INITIAL  RESULTS 


400 


c 


g  200 
E 

S  o 
t 

c-200 

o 

N 


observations  . 


•  »  .   •   •  . 


The  spatial  error  associated  with  the  points  from  Sourcel  is  illustrated  in 
figure  2.  Extreme  values  are  not  shown.  Source2  has  similar  errors. 

Figure  2.  Scatter  diagram  of 

Sourcel  :  magnitude  and  direction  of  errors  in  x  and  y  the  horizontal  error  in  X  and  y 

observed  in  quadrat  locations  from 
sourcel.  The  diagram  indicates 
that  errors  in  x  and  y  tend  to  be 
negatively  correlated  suggesting  an 
anisotropic  error  distribution.  The 
mechanisms  behind  this  are 
unknown. 

The  importance  of  the 
spatial  error  in  terms  of 
inducing  errors  in  the  estimates 
of  elevation  and  slope  is 
illustrated  by  figure  3,  which 
shows  deviations  from  the 
correct  values  of  slope  and 
elevation.  Figure  3  clearly  illustrates  that  for  a  given  degree  of  spatial  error  the 
implications  of  the  error  are  greater  for  estimates  of  slope  than  for  elevation. 

The  results  given  in  figure  3  can  be  summarised  using  values  for  slope  and 
elevation  between  the  true  locations  and  the  mapped  locations  as  follows: 


^oo 


~2txr 


horizontal  error  (metres)  in  x 


Slope  (percent  rise) 

Elevation  (metres) 

Point  locations  from 
Sourcel 

0.181 

0.952 

Point  locations  from 
Sourcel 

0.318 

0.976 

MODELLING  THE  IMPORTANCE  OF  SPATIAL  ERRORS 


Spatial  error  is  present  when  xq  *  xi ,  xq  being  the  true  place  of  observation, 
and  x\  the  location  mapped.  Let/(x)  be  a  function  denoting  the  probability  density 
of  the  event  x  =  x\  and  Z(xi)  be  the  actual  observation  (of  slope,  insolation,  soil 
depth),  while  Z(xq)  is  the  correct  observation  (unknown). 

The  question  then,  is:  when  xq  is  estimated  by  xi,  what  is  the  expected 
outcome  of  observations  of  some  spatial  variable,  Z(xi ). 

We  can  state  the  following: 

E[Z(Xl)]   =  \Z(x)f(x)dx  =  JZ(xo  +  h)/(xo  +  h)<ai  (1) 
where  h  is  a  vector  such  that  x  =  xq  +  h 

This  is  the  same  form  of  equation  identified  by  Allen  and  Starr  (1982)  as 
corresponding  to  observation  of  Z(x)  at  a  coarser  scale,  with  the  scaling  function 
weights  being  determined  by/(xQ  +  h).  From  this  we  can  say  that  spatial  error  in 
point  locations  has  similar  consequences  for  the  expected  value  of  observation  as 
would  viewing  at  a  broader  scale. 


579 


Figure  3.  Plots  of  estimates  of  elevation  and  slope  from  quadrat  locations  from  sourcel 
and  sourcely  against  the  correct  estimate.  Departures  from  'y=x'  indicate  errors  in  the 
estimate  due  to  quadrat  location  errors.  Relatively  large  errors  are  introduced  into 
estimates  of  slope  cf.  elevation,  irrespective  of  the  source  of  the  quadrat  error  {sourcel  vs. 
sourcel).  The  type  of  variable  observed,  rather  than  the  specific  amount  of  spatial  error  in 
the  quadrats,  appears  to  dictate  whether  the  errors  introduced  will  be  substantial  or 
negligible. 

In  qualitative  terms,  the  results  become  fuzzy.  Statistically,  the  correlation 
between  the  observed  value  Z(x\)  and  the  true  value  Z(xq)  will  drop  below  1. 
unless  the  spatial  error  is  zero.  The  severity  of  the  error  will  be  indicated  by  the  R* 
value  between  the  observed  value,  Z(\\)  and  the  true  value  Z(xq). 

For  a  known  spatial  error,  h  the  co variance,  cov(h),  between  the  imperfect 
observation,  taken  at  location  xi  =  xq  +  h,  and  the  true  observation,  taken  at  xq,  is 

cov(h)  =  E[(Z(xo+h)  -  HzXZ&6>  -  |iz)]  (2) 
The  correlation  coefficient  between  Z(x\)  and  Z(xq)  for  a  known  error  h  is 
given  by 

R(h)2  =  (cov(h)  I  a2)2  (3) 
where  /?(h)2  is  commonly  interpreted  as  the  proportion  of  the  variance  in 
Z(xq)  accounted  for  by  Z(x\). 


580 


The  covariance  is  related  to  the  semivariance  by  7(h)  =  o2  -  cov(h),  (assuming 
that  Z  is  a  stationary  function  and  that  the  error  model  is  isotropic)  so  equation  3 
can  be  re-stated  as 

[/?(h)]2  =  (l  -7(h) /a2)2  (4) 
To  progress  from  the  simple  case  of  known  error  h  to  a  probability  model  of 
error /(h)  we  calculate  the  expected  value  of  cov(h)  using  simply: 

E[R2]  =  J/(h)  R  (h)2  dh 

=  j/(h)(l-7(h)/rj2)2Jh  (5) 
This  gives  a  direct  mathematical  expression  to  estimate  the  importance  of 
spatial  error  in  a  given  situation,  requiring  only  a  model  of  the  spatial  error  of  the 
point  observations  and  a  correlogram  or  variogram  of  the  spatial  variable  Z.  The 
semivariance,  7(h)  can  be  thought  of  as  that  part  of  the  total  variance  which  is 
introduced  at  lag  h. 

Equation  5  also  illustrates  the  relationship  between  the  magnitude  of  the 
spatial  error  and  the  inherent  scaling  of  the  variable  in  determining  the  importance 
of  the  spatial  error,  discussed  earlier.  Clearly  if  /(h)  is  high  only  where  7(h)  is  near 
0,  the  error  is  less  important  than  if /(h)  is  high  over  a  wide  range  of  7(h)  including 
where  7(h)  ~  a2.  Forms  of  /(h)  which  either  have  a  non-zero  mean  error,  E[(h)]  * 
0,  or  which  are  multi-modal,  with  a  peak  some  distance  from  xq  (eg.  Bolstad  et  al. 

1990)  will  tend  to  introduce  more  significant  errors. 

Prediction  of  the  errors  in  elevation  and  slope  introduced  through  spatial 

error  in  coordinate  locations. 

Equation  5  demonstrates  that  information  on  the  distribution  of  spatial  errors 
in  the  point  data  and  a  knowledge  of  the  variogram  of  Z  are  required  to  estimate 

giving  a  direct  assessment  of  the  data  quality  of  the  points,  for  observation  of 
a  given  variable  Z. 

If  isotropic  variograms  and  error  models  are  assumed,  equation  5  simplifies  to 

E[R2]  =  lf(h)R(h)2dh  =    \f{h)  ( 1  -  y{  h)  I  o2)2  dh  (6) 
where  h  =  Ihl. 

In  the  following  section  this  method  is  applied  to  the  datasets  used  here.  The 
Excel  software  was  used  to  manipulate  data  and  to  apply  mathematical  formulae, 
while  variograms  were  estimated  using  software  written  by  the  author.  The 
continuous  functions  indicated  by  equation  6  were  approximated  using  discrete 
forms  with  an  interval  of  40  metres  (/z=0,40,  ....8000). 

Models  of  spatial  error 

An  isometric  error  model  was  adopted.  Normal  and  lognormal  distributions 
were  fitted  to  observed  errors  using  the  method  of  moments  (figure  4). 

The  parametric  models  enable  the  standard  deviation  to  be  used  as  a 
parameter  to  vary  the  model,  allowing  sensitivity  testing.  The  assumption  of  an 
isometric  normal  model,  f(h),  has  the  further  advantage  of  requiring  only  one 
parameter,  ae.  The  convention  of  epsilon  =  2o e  is  adopted,  thus  in  terms  of  an 


581 


epsilon  model  of  error  we  can  state  that  95%  of  the  data  points  are  expected  to  lie 
within  epsilon  metres  of  the  mapped  point. 


horizontal  error  (metres)  horizontal  error  (metres) 


Figure  4.  Hots  of  the  cumulative  distribution  of  horizontal  error  for  source  1  quadrats 
(left)  and  soured  quadrats  (right);  and  normal  and  lognormal  probability  models  fitted 
using  the  method  of  moments.  An  isometric  error  distribution  is  assumed.  For  sourcel,  the 
normal  distribution  was  fitted  after  exclusion  of  extreme  values.  Neither  probability  model 
deals  with  sourcel  very  well,  however  the  normal  distribution  was  chosen  in  preference  to 
the  lognormal. 

ASSESSING  DATA  QUALITY  FOR  SLOPE  AND  ELEVATION 

The  discrete  version  of  equation  6  was  used  to  calculate  the  expected  R2  value 
for  true  and  estimated  values  of  slope  and  elevation  from  sourcel  and  sourcel  for 
the  values  of  epsilon  given  in  Table  1 . 

E[R2]  =    X    [F(h+A/2)  -  F(h-A/2)][1  -yity/o1)]2  + 

h=A,  2A, ...  ,8000 

[F(A/2)-F(0)][l-7(0)/o2)]2  (7) 
where:  F(h)  is  the  cumulative  normal  probability  distribution  with  standard 
deviation  of  epsilon/2,  truncated  to  exclude  negative  values  (F(0)  =  0)  and  scaled 
to  maintain  F(<»)  =  1 ;  h  is  the  magnitude  of  the  spatial  error,  and  the  lag  distance 
for  y(h);  A  is  an  interval  (40  m),  cr2  is  the  observed  variance  of  Z  (slope  or 
elevation,  as  appropriate). 

Using  equation  7  the  predicted  R2  values  for  the  observations  of  slope  and 
elevation,  from  sourcel  and  source2,  were  calculated  and  compared  with  the 
observed  values  (table  1). 


Table  1.  Modelled  and  observed  (in  brackets)  correlations  (R2)  between  observations 
of  elevation  and  slope  from  true  data  point  locations  and  data  point  locations  containing 
spatial  error.  Epsilon  =  2a.  ===___^_______________  


Slope  (percent  rise) 

Elevation  (metres) 

Point  locations  from  Sourcel. 

Epsilon  =  384 

0.303  (0.181) 

0.957  (0.952) 

Point    locations    from  Soured. 

Epsilon  =  320 

0.343  (0.318) 

0.967  (0.976) 

582 


In  most  practical 
situations,  y(h)  is  readily 
estimated  for  datasets  in 
a  GIS,  but  the  precise 
nature  of  F(h)  will  be 
unknown.  To  develop 
some  idea  of  the 
sensitivity  of  data 
quality  to  /(/*),  E[R2] 
was  calculated  for  a 
wide  range  of  epsilon. 
Results  are  shown  in 
figure  5,  from  which  it  is 
clear  that  the  modelled 
R2  values  for  elevation 
from  source  1  and 
source!  are  closely 
predicted.  R2  for  slope 
is  accurately  predicted 
for  source2,  but  for 
sourcel  is  optimistic, 
reflecting  the  extreme 
values  not  catered  for  by 
the  normal  distribution 
(figure  4). 

Figure  5.  Predictions 

of  R2  value  for  a  wide 
range  of  error  bands, 
compared   with  observed 

R2  values. 

The  contrast 
between  figures  5  and  6 
shows  that  the  key 
factor  determining  the  R2  between  correct  and  incorrect  values  of  elevation  and 
slope  is  not  the  spatial  error  in  the  observation  points,  but  the  inherent  spatial 
structure  of  the  variables  being  observed. 

Thus,  even  the  very  coarse  models  of  spatial  error  used  here  are  sufficient  to 
quite  accurately  model  the  quality  of  the  quadrat  data  sources  for  overlay  with 
other  spatial  data  themes. 

DISCUSSION 

These  results  have  wide-ranging  practical  value.  They  suggest  that  the  detail 
of  the  probability  error  model  /(x),  is  less  important  in  determining  data  quality 
than  the  inherent  properties  of  the  surfaces  being  interrogated.  Only  the  main 
features  of/(x)  seem  to  be  important. 

The  results  also  suggest  an  approach  to  management  of  GIS  error  which 
integrates  spatial  and  attribute  error.  If  the  value  of  a  variable  is  estimated  from 
interrogation  of  a  data-surface  Z,  where  Z  is  a  model  with  residual  (attribute)  error 
o2ez  ,  the  variance  of  Z,  at  any  given  place,  due  to  spatial  and  attribute  error  can 


583 


be  calculated  as  the  expected  semivariance,  plus  the  variance  of  the  residual  error 
of  the  model. 

°2Z  =  E[yz]  +  G2ez    where  E^  =  -f/z(h)  Yz*)  (8) 
In  which /7(h)  is  the  pdf  of  the  spatial  errors  in  variable  Z,  while  Yz(h)  is  the 

semivariance  of  variable  Z  readily  estimated  readily  from  the  data-surface  Z. 

Applying  equation  8  to  the  DEM  used  here,  assuming  isometric  normally 

distributed  spatial  errors  with  epsilon  =  25,  and  an  RMS  of  5  (G2ez  =  25)  gives 

E[Yt]  ~  3.7,  a2^  =  3.7  +  25.  Thus  (for  elevation)  the  influence  of  spatial  error  is 

small  compared  with  residual  error. 

REFERENCES 

Allen  T  F  H,  Starr,  T  B  (1982)  Hierarchy:  perspectives  for  ecological  complexity. 

University  of  Chicago  Press 
Bolstad,  P  V,  Gessler,  P,  Lillesand,  T  M  (1990)  Positional  uncertainty  in  manually 

digitised  map  data.  International  Journal  of  Geographic  Information  Systems.  4 

399 

Burrough,  P  A.  (1986)  Principles  of  Geographical  information  systems  for  land 

resources  assessment.  Clarendon  Press.  Oxford. 
Goodchild,  M  F.  The  issue  of  accuracy  in  global  databases.  (1988)  In  Mounsey,  H, 

Tomlinson,  R  F.  (eds.)  Building  databases  for  global  science.  Proceedings  of 

the  first  meeting  of  the  International  Geographical  Union  Global  Database 

Planning  Project.  Tylney  Hall,  Hampshire,  UK,  May  1988.  Taylor  and  Francis. 

London  &  New  York,  1988 
Goodchild,  M  F,  Guoqing,  S,  Shiren,  Y.  (1992)  Development  and  test  of  an  error 

model  for  categorical  data.  International  Journal  of  Geographical  Information 

Systems.  6  87-104 

Hutchinson,  M  F  (1988)  Calculation  of  hydrologically  sound  digital  elevation 

models.  Proceedings  of  the  Third  International  Symposium  on  Spatial  Data 

Handling.  Sydney,  Australia. 
Hutchinson,  M  F  (1989)  A  new  procedure  for  gridding  elevation  and  stream  line 

data  with  automatic  removal  of  spurious  pits.  Journal  of  Hydrology  106  211- 

232 

Turner,  L  A,  Mueck,  S  G.  (1992)  The  vegetation  of  the  Sardine,  Rich  and  Ellery 
forest  blocks,  Orbost  region,  Victoria.  Silvicultural  Systems  Project  Technical 
Report  number  9.  Department  of  Conservation  and  Environment,  Melbourne, 
Australia. 

BIOGRAPHICAL  SKETCH 

Adam  Lewis  is  the  Senior  GIS  Scientist  in  the  Natural  Resource  Systems 
Branch  (NRS)  of  the  Department  of  Conservation  and  Natural  Resources,  based  in 
Melbourne,  Australia.  He  has  experience  in  Forest  and  Land  Management,  and 
recently  completed  a  PhD  and  the  Australian  National  University. 

Michael  Hutchinson  is  a  Senior  Fellow  at  the  Australian  National  University 
whose  primary  interest  is  the  spatial  and  temporal  analysis  of  physical 
environmental  data. 


584 


Moving  Into  Secondary  Map 
Projections  - 
An  Analysis  of  Potential 
Inconsistencies  In  Spatial  Data 

Mohamad  N.  Said1  and  Peter  F.  Fisher2 


Abstract. — Implementation  of  GIS  functions  especially  those  involving 
complex  mathematical  computation  such  as  map  projection  may  result  in 
a  certain  amount  of  deformation  in  spatial  data.  In  this  research,  several 
experiments  using  a  test  pattern  are  carried  out  to  test  the  map  projection 
function  provided  by  two  selected  commercial  GIS's.  Experiments  include 
the  change  of  projection  within  the  same  and  across  categories.  It  is  found 
that  there  are  geometrical  and  topological  inconsistencies  as  a  results  of 
these  transformations.  Most  are  relatively  small  and  probably  the  result  of 
digital  rounding  errors  through  the  computational  process. 


INTRODUCTION 

All  spatial  data  layers  stored  in  a  Geographical  Information  Systems  (GIS)  are 
based  on  a  certain  type  of  map  projection  which  defines  an  x-y  Cartesian 
coordinate  system  for  all  features  contained  in  the  database.  Three  categories  of 
map  projection  are  commonly  identified:  conformal  projections  maintain  the  shape; 
equal  area  projections  maintain  the  area;  and  equidistant  projections  preserve  the 
distance. 

Having  a  multiple-theme  database,  therefore,  requires  all  themes  to  be  in  a 
common  projection,  and  if  they  are  not  some  number  will  need  to  be  transformed 
to  fit  a  selected  one.  Such  transformations  are  common  in  practice  and  may  take 
place  a  number  of  times,  especially  when  GIS  users  exchange  digital  map  data  or 
require  different  properties  in  their  output  maps.  The  reprojection  may  involve  a 
change  of  parameters  contained  in  the  output  projection  including  the  geodetic 
datum,  the  central  meridian,  the  standard  parallels,  the  origin  etc. 


1  Research  student.  Dept.  of  Geography,  University  of  Leicester,  U.K. 

2  Senior  lecturer,  Dept.  of  Geography,  University  of  Leicester,  U.K. 


585 


It  is  anticipated  that  the  reprojection  of  original  maps  into  secondary  projections 
may  cause  a  certain  amount  of  deformation  and  the  magnitude  and  pattern  of  the 
errors  may  be  different  when  transformations  are  done  using  different  GIS 
packages.  This  paper  reports  the  outcome  of  the  research  being  carried  out  to 
demonstrate  these  potential  inconsistencies  and  discrepancies. 

RELATED  WORKS 

Some  works  on  benchmarking  of  GIS  functions  or  algorithms  do  exist  in  the 
literature.  Piwowar  et  al.  (1990)  reported  the  research  outcome  of  raster-to-vector 
conversion  using  three  different  algorithms  while  Knaap  (1992)  took  further  tests 
of  the  same  work  by  comparing  the  results  output  by  eight  different  GIS  packages. 
Knaap  showed  that  no  two  of  the  output  produced  the  same  rasterization  of  the 
test  vector  patterns,  although  there  were  some  reassuring  consistencies.  Fisher 
(1993)  explored  the  viewshed  functions  and  reported  the  extent  of  variability  in  the 
viewable  area  when  the  function  is  implemented  in  different  packages.  He  further 
highlighted  the  need  for  standards  and/or  empirical  benchmark  datasets  for  GIS 
functions  and  the  desirability  of  publications  of  algorithms  used  in  GIS  functions. 

Franklin  (1984)  demonstrated  the  topological  effects  of  changing  scale  and 
rotation  when  rounding  occurs  in  integer  arithmetic.  Burrough  (1986)  quoted  the 
work  by  Gruenberger  ( 1 984)  of  how  computer  word  length  can  have  a  cumulative 
effect  after  a  real  number  is  repeatedly  squared  on  different  platforms. 

Masters  (1992)  stressed  the  importance  of  spatial  data  accuracy  which  has  to  be 
considered  in  the  context  of  coordinate  systems  defined  by  geodetic  datums  and 
map  projections.  He  further  made  a  comment  on  the  continuing  difficulties  in  GIS 
projects  due  to  the  assumption  made  by  many  GIS's  that  the  coordinate  systems 
that  maps  are  based  on  are  all  the  same  and  all  have  the  same  accuracy. 

EXPERIMENTAL  DESIGN  AND  TEST  DATA 

The  Map  Projections 

In  the  implementations  tested  in  this  study,  three  projections  are  selected.  These 
are  Lambert  Conformal  Conic  (or  Lambert),  Universal  Transverse  Mercator  (or 
UTM)  and  Albers  Conic  Equal  Area  (or  Albers)  projections.  The  parameters  used 
throughout  the  experiments  are  as  follows: 

Lambert  Conformal  Conic 

Spheroid  Clarke  1866 

Units  meter 

Central  Meridian  75°  00"  00'  W 


586 


Standard  Parallels 

Latitude  of  Projection  Origin 
False  Easting  (m) 
False  Northing  (m) 


2°  20'  00"N 
3°  40'  00"N 
0°  00'  00" 
0.0000 
0.0000 


Universal  Transverse  Mercator  (UTM) 


Spheroid 

Units 

Zone 


Clarke  1866 

meter 

18 


Albers  Equal  Area 

Spheroid 

Units 

Central  Meridian 
Standard  Parallels 

Latitude  of  Projection  Origin 
False  Easting  (m) 
False  Northing  (m) 


Clarke  1866 
meter 

75°  00"  00'  W 
2°  20'  00"N 
3°  40'  00"N 
0°  00'  00" 
0.0000 
0.0000 


The  Test  Data 

An  experimental  test  pattern  of  16  square  polygons  (figure  1)  covering  2°  by  2° 
in  the  equatorial  region  is  designed  to  fit  the  Zone  18  band  of  the  United  States 
(longitudes:  74°W  to  76°W).  The  lower  limit  of  the  latitude  is  at  2°  00'  00"N  while 
the  upper  limit  is  at  4°  00'  00"N.  The  Central  Meridian  of  this  Zone  is  chosen  to 
be  at  the  longitude  of  75°W  which  is  along  the  central  line  of  the  quadrangle. 

The  GIS  Packages 

Two  commercial  GIS  packages  are  used  in  the  study,  namely  Arc/Info  (version 
7.0.3)  and  Genamap  (version  5.1).  Arc/Info  runs  on  an  IRIX  mainframe  while 
Genamap  does  on  an  IBM  RISC6000  workstation.  The  original  test  data  are 
however  created  and  keyed  independently  to  avoid  any  deformation  in  the  data 
transfer  process. 

The  Transformation  Schemes 

Experiments  are  carried  out  to  test  for  any  inconsistencies  in  the  conformal  map 
(Lambert)  when  it  is  transformed  into  other  projections  within  its  category  as  well 


587 


as  across  the  other  categories.  The  transformations  are  accomplished  according  to 
the  following  schemes: 


1+ 


761  ► 


27* 


Zone  18 


2nd  st  indard  pars  llel :  3  0  40' 


It" 


5-h 


91- 


21- 


6  + 


lOf 


141- 


lst sU  ndard_para lleU  2J_20^ 


-0O 


|  7f 

•c 

*hI  -H- 


2 

1  111- 

'J 

$4  ^4 


151- 


3«4000'  N 


00"  N 
4f 


121- 


161- 
|00"N 


-00- 


3° 30' N 


f£<>3000'  N 


2°  30'  N 


2°  00'  N 


76°00'W    75°30'W    75°00' W    74°30' W  74°00'W 

16%  node  ID 
161"  polygon  ID 


Figure  1.  —  The  test  pattern 


a)  Transformation  of  geographical  coordinates  into  Lambert  Conformal  Conic 

GEOG         — >  LMB1 

GEOG  represents  the  original  map  in  geographical  coordinates  (latitude  and 
longitude)  which  is  transformed  into  Lambert  Conformal  Conic  projection 
producing  LMB1. 

b)  Transformation  Within  Conformal  Category 

LMB1  — ->      UTM        --->  LMB2 

The  initial  map  in  Lambert  projection  as  obtained  in  (a)  is  further  transformed 
into  another  conformal  projection,  in  this  test  UTM  is  chosen.  This  map  is  then 
reprojected  back  into  Lambert  Conformal  Conic  producing  LMB2  . 

c)  Transformation  Across  Category 

LMB1  --->      ALBERS      ->  LMB3 


588 


In  this  case  the  interchange  projection  is  equal  area  (ALBERS).  This  produces 
the  third  Lambert  map  named  as  LMB3. 

The  purpose  of  reprojection  as  carried  out  in  (b)  and  (c)  above  is  to  enable 
checking  of  inconsistency  between  LMB1  and  LMB2. 

Further  experiments  are  also  made  by  overlaying  maps  of  similar  projection 
before  and  after  transforming.  This  includes  a  union  process  of  combining  LMB1A 
and  LMBIg  as  well  as  LMB2A  and  LMB2G  (the  subscripts  A  and  G  denote  maps 
transformed  with  Arc/Info  and  Genamap  GIS's  respectively).  Any  new  intersection 
points  will  indicate  the  occurrence  of  topological  errors. 

Table  1. — Internal  consistency  check  -  shifts  in  coordinates.  

Node  Arc/Info  Genamap 

ID  LMB2  -  LMB1  LMB2  -  LMB1 


diff  in  x  (m)  diff  in  y  (m)  diff  in  x  (m)       diff  in  y  (mi 


1 

-0.0070 

0.0000 

0.0000 

0.0000 

2 

0.0070 

0.0000 

0.0000 

0.0000 

3 

-0.0150 

0.0000 

0.0000 

0.0000 

4 

0.0150 

0.0000 

0.0000 

0.0000 

5 

0.0000 

0.0000 

0.0000 

0.0000 

6 

0.0000 

0.0000 

0.0000 

0.0000 

7 

-0.0320 

0.0000 

0.0000 

0.0000 

8 

-0.0040 

0.0000 

0.0000 

0.0000 

9 

-0.0280 

0.0000 

0.0000 

0.0000 

10 

0.0000 

0.0000 

0.0000 

0.0000 

11 

0.0080 

0.0000 

0.0000 

0.0000 

12 

0.0230 

0.0000 

0.0000 

0.0000 

13 

-0.0040 

0.0000 

0.0000 

0.0000 

14 

0.0040 

0.0000 

0.0000 

0.0000 

15 

0.0000 

0.0000 

0.0000 

0.0000 

16 

-0.0150 

0.0000 

0.0000 

0.0000 

17 

-0.0160 

0.0000 

0.0000 

0.0000 

18 

0.0080 

0.0000 

0.0000 

0.0000 

19 

-0.0080 

0.0000 

0.0000 

0.0000 

20 

0.0000 

0.0000 

0.0000 

0.0000 

21 

-0.0150 

0.0000 

0.0000 

0.0000 

22 

-0.0160 

0.0000 

0.0000 

0.0000 

23 

0.0080 

0.0000 

0.0000 

0.0000 

24 

-0.0080 

0.0000 

0.0000 

0.0000 

25 

0.0000 

0.0000 

0.0000 

0.0000 

589 


RESULTS 


Analyses  are  carried  out  on  the  resulting  Lambert  maps,  represented  in  bold  in 
previous  section.  Observations  are  made  on  any  shirts  in  the  coordinates  of  nodes 
as  well  as  in  the  area  of  polygons  contained  in  the  maps. 

The  internal  consistency  of  each  software  is  checked  by  comparing  the  first  and 
the  second  Lambert  maps  (see  transformation  schemes  (b)  and  (c)).  From  table  1, 
it  can  be  seen  that  transformation  between  conformal  projections  with  Arc/Info 
produces  some  amount  of  error  in  the  feature  coordinates.  These  errors  occur  only 
in  the  x  direction  and  are  very  small,  with  a  maximum  magnitude  of  0.032m, 
relative  to  the  coverage  of  the  study  pattern.  As  a  consequence  of  these  shifts,  the 
polygons  suffer  a  change  in  area  between  256  to  1280  square  meters.  Genamap's 
output  seems  to  be  more  consistent  for  this  particular  test,  where  no  errors  are 
shown.  Transformation  via  the  equal  area  projection  however  does  not  cause 
any  error  in  either  package. 


Table  2. — Comparison  of  absolute  coordinates:  Arc/Info  vs  Genamap 


Node 

T  A  M  D  t 

LMBIg  - 

LMBIa  (m) 

LMd2g  - 

LMB2a  (m) 

LMB3a  (m) 

ID 

Diff.  in  x 

Diff.  in  y 

Diff.  in  x 

Diff.  in  y 

Diff.  in  x 

Diff.  in  y 

1 

-0.0011 

0.0128 

0.0059 

0.0128 

-0.0011 

0.0128 

2 

0.0011 

0.0128 

-0.0059 

0.0128 

0.0011 

0.0128 

3 

0.0017 

0.0014 

0.0167 

0.0014 

0.0017 

0.0014 

4 

-0.0017 

0.0014 

-0.0167 

0.0014 

-0.0017 

0.0014 

5 

0.0000 

0.0073 

0.0000 

0.0073 

0.0000 

0.0073 

6 

-0.0035 

-0.0110 

-0.0035 

-0.0110 

-0.0035 

-0.0110 

7 

0.0035 

-0.0110 

0.0355 

-0.0110 

0.0035 

-0.0110 

8 

0.0005 

-0.0086 

0.0045 

-0.0086 

0.0005 

-0.0086 

9 

-0.0005 

-0.0086 

0.0275 

-0.0086 

-0.0005 

-0.0086 

10 

0.0000 

-0.0085 

0.0000 

-0.0085 

0.0000 

-0.0085 

11 

-0.0015 

-0.0053 

-0.0095 

-0.0053 

-0.0015 

-0.0053 

12 

0.0015 

-0.0053 

-0.0215 

-0.0053 

0.0015 

-0.0053 

13 

0.0015 

0.0107 

0.0055 

0.0107 

0.0015 

0.0107 

14 

-0.0015 

0.0107 

-0.0055 

0.0107 

-0.0015 

0.0107 

15 

0.0000 

0.0061 

0.0000 

0.0061 

0.0000 

0.0061 

16 

0.0010 

0.0120 

0.0160 

0.0120 

0.0010 

0.0120 

17 

-0.0010 

0.0120 

0.0150 

0.0120 

-0.0010 

0.0120 

18 

-0.0013 

0.0107 

-0.0093 

0.0107 

-0.0013 

0.0107 

19 

0.0013 

0.0107 

0.0093 

0.0107 

0.0013 

0.0107 

20 

0.0000 

0.0003 

0.0000 

0.0003 

0.0000 

0.0003 

21 

0.0000 

-0.0062 

0.0150 

-0.0062 

0.0000 

-0.0062 

22 

0.0000 

-0.0062 

0.0160 

-0.0062 

0.0000 

-0.0062 

23 

0.0022 

0.0063 

-0.0058 

0.0063 

0.0022 

0.0063 

24 

-0.0022 

0.0063 

0.0058 

0.0063 

-0.0022 

0.0063 

25 

0.0000 

0.0051 

0.0000 

0.0051 

0.0000 

0.0051 

590 


A  comparison  of  absolute  coordinates  is  shown  in  table  2.  There  is  a  certain 
amount  of  discrepancy  in  both  x  and  y  coordinates.  The  pattern  of  these 
differences,  in  all  three  comparisons,  is  found  to  be  symmetrical  about  the  central 
meridian.  These  discrepancies  have  caused  the  change  in  area  of  the  polygons 
between  10  to  1024  square  meters.  The  small  differences  seen  above  are  sufficient 
to  cause  line  intersections  when  maps  from  different  packages  are  combined.  For 
example,  by  combining  LMB1A  and  LMB1G  ,  there  are  68  polygons  created 
instead  of  16.  This  means  52  sliver  polygons  have  been  generated  even  though  they 
are  small,  i.e.  with  a  maximum  area  of  about  638  square  meters.  It  is  worse  when 
LMB2A  and  LMB2g  are  combined,  where  72  polygons  are  generated  giving  an 
extra  of  56  with  a  maximum  area  of  about  939  square  meters.  Graphically,  these 
polygons  are  invisible  without  considerable  magnification  but  each  of  them  adds  up 
one  new  records  in  the  attribute  database.  Thus  in  the  first  example,  52  new 
records  occur  in  the  attribute  database. 

CONCLUSIONS 

From  the  various  tests  described  and  implemented  throughout  this  research,  it 
has  been  shown  that  errors  do  exist  as  a  results  of  moving  from  one  projection  to 
another.  They  are  found  to  be  very  small  and  may  be  due  to  digital  rounding.  It  is 
believed  that  different  GIS  packages  may  or  may  not  introduce  internal 
inconsistencies  but  they  may  produce  different  results  when  compared  to  each 
other.  Further  experiments  such  as  transformations  of  maps  at  varying  latitudes 
and  with  other  projection  combinations  have  been  examined. 

Integrating  datasets  using  derived  maps,  in  this  case  maps  with  secondary 
projections,  from  different  packages  may  cause  spatial  problems  such  as  line 
intersections  and  sliver  polygons.  This  problem  is  inherent  in  the  maintenance  of 
the  spatial  and  attribute  databases  when  new  unwanted  features  are  generated.  It 
has  been  shown  that  with  small  discrepancies  in  feature  coordinates,  the  database 
has  been  grown  by  three  times. 

As  a  final  conclusion,  further  benchmarking  tests  on  map  projection  function 
with  other  GIS  packages  will  add  to  the  knowledge  of  potential  inconsistencies  in 
the  spatial  data.  This  will  reassure  the  GIS  users  about  the  function  they  are  using. 

REFERENCES 

Briggs,  D.  and  Mounsey,  H.,  1989.  Integrating  Land  Resource  Data  Into  A  European 
Geographical  Information  System  :  Practicalities  and  Problems.  Applied  Geography, 
Vol.  9(1989).  pp  5-20. 

Burrough,  P.  A.  1986.  Principles  of  Geographical  Information  Systems  For  Land 
Resources  Assessment.  Clarendon  Press:  Oxford.  194p. 


591 


Fisher,  P.  F.  1993.  Algorithm  and  implementation  Uncertainty  In  Viewshed  Analysis. 
International  Journal  of  Geographical  Information  Systems,  Vol.  7,  No.  4.  London: 
Taylor  and  Francis,  pp  331-347. 

Franklin,  R.,  1984.  Cartographic  Errors  Symptomatic  of  Underlying  Algebra  Problems  - 
Proceedings  of  the  International  Symposium  On  Spatial  Data  Handling,  1994, 
Zurich,  pp  190-208. 

Knaap,  W.G.  M.  VAN  DER,  1992.  The  Vector  To  Raster  Conversion:  (mis)use  In 
Geographical  Information  Systems.  International  Journal  of  Geographical  Information 
Systems,  Vol.  6,  No.  2.  London:  Taylor  and  Francis,  pp  159-170. 

Maling,  D.H.,  1991.  Coordinate  Systems  And  Map  Projections  For  GIS  -  in 
Geographical  Information  Systems:  Principles  And  Applications,  Vol.  1  (ed. 
Maguire,  D.J.,  Goodchild,  M.F.  and  Rhind,  D.W).  Longman  Group  UK  Ltd. 

Masters,  M.,  1992.  Digitization  and  Database  Accuracy.  Proceedings  of  GIS/LIS  '92 
Annual  Conference  and  Exposition,  San  Jose,  California,  pp  522-531. 

Piwowar,  J.M.,  LeDrew,  E.F.,  and  Dudycha,  D.J. ,  1990.  Integration  of  Spatial  Data  In 
Vector  and  Raster  Formats  In  A  Geographic  Information  System  Environment. 
International  Journal  of  Geographical  Information  Systems,  Vol.  4,  No.  4.  London: 
Taylor  and  Francis,  pp  429-444. 

Snyder,  J.P.  1987  Map  Projections  -  A  Working  Manual.  U.S.  Geological  Survey 
Professional  Paper  1395.  p383. 

BIOGRAPHICAL  SKETCH 

Mohamad  N.  Said  is  currently  a  research  student,  doing  his  PhD  in  GIS  with  particular 
interest  in  studying  the  positional  errors  in  spatial  data  due  to  the  change  of  map  projection 
systems.  He  obtained  his  bachelor  degree  in  Surveying  and  Mapping  Science  from  the 
North  East  London  Polytechnic,  U.K.  (1985)  and  master  degree  from  Curtin  University  of 
Technology,  Australia  (1990). 

Peter  F.  Fisher  is  a  senior  lecturer  at  the  Department  of  Geography,  University  of 
Leicester  with  special  interests  in  error  modelling,  propagation  and  visualization.  He  is 
currently  the  editor  of  International  Journal  of  Geographic  Information  Systems.  He  holds 
a  B.Sc.  in  Environmental  Sciences  from  the  Lancaster  University  (1977),  an  M.Sc.  in 
Pedology  and  Soil  Survey  from  Reading  University  and  a  PhD  from  Kingston  Polytechnic 
(1982). 


592 


Filling  in  Missing  Forestry  Data: 
Exploring  Autocorrelational  Techniques 

Alissa  N.  Antle1  and  Peter  L.  Marshall2 


Abstract. — A  crucial  component  of  any  effective  approach  to 
long-term  management  and  protection  of  valuable  forest  resources 
is  the  ability  to  represent  and  model  spatial  forest  ecosystem 
processes.  Geographic  information  systems  (GIS)  integrated  with 
spatial  forest  models  provide  an  appropriate  spatial  framework  for 
forest  ecosystem  modeling.  In  order  to  be  effective,  accurate 
ecological  data  must  be  used  as  input  to  GIS  and  spatial  forest 
models.  It  is  often  the  case  that  the  forest  inventory  attribute  of 
interest  may  have  incomplete  coverage  for  the  area  under 
management.  In  these  cases  it  is  mandatory  to  have  techniques  that 
can  predict  the  missing  values  while  preserving  the  complex  spatial 
patterns  between  forest  inventory  attributes.  One  of  the  most 
prominent  spatial  patterns  found  in  ecological  data  is  the  presence 
autocorrelation. This  paper  introduces  several  autocorrelational  tests, 
describes  different  methods  of  representing  spatial  autocorrelational 
structures,  and  discusses  various  approaches  which,  either  explicitly 
or  implicitly,  use  this  information  to  predict  missing  values  in 
ecological  data  sets. 


INTRODUCTION 

In  order  for  spatial  forest  models  to  be  effective,  accurate  and  complete 
ecological  data  must  be  used.  Spatial  data  quality  has  received  much  recent 
attention  in  the  GIS  community  (Goodchild  and  Gopal  1989,  Chrisman  1991, 
Guptill  and  Morrison  1995).  The  term  completeness  has  been  used  in  reference  to 
both  resolution  and  missing  values.  Missing  values  are  those  which  are  expected 
to  appear  in  the  data  set  but  do  not.  In  forest  inventory  data  sets,  values  may  be 
missing  due  to  collection  errors  or  omissions.  In  particular,  attributes  required  for 
a  specific  forest  model  may  have  not  been  collected,  or  because  of  the  expense  or 
effort  required  to  collect  the  data,  only  a  limited  set  of  data  may  be  available.  In 
order  to  ensure  that  data  sets  are  fit  for  forest  modeling  purposes,  techniques  that 
can  predict  missing  values  are  necessary  to  complete  data  sets  prior  to  use  in 


Department  of  Geography,  University  of  British  Columbia.  Canada 
Department  of  Forest  Resources  Management.  University  of  British  Columbia,  Canada 


593 


existing  models  or  in  developing  new  models. 

While  conventional  statistical  techniques  have  been  widely  used  to  predict 
missing  values,  most  of  these  techniques  substitute  mean  or  median  values  for 
missing  data.  These  approached  do  not  preserve  the  complex  spatial  patterns  which 
may  be  found  among  multivariate  forest  inventory  attributes.  In  addition,  most 
interpolative  approaches  are  based  on  the  assumption  of  spatially  independent  data. 
Simply  stated,  the  problem  is  how  to  best  predict  missing,  autocorrelated  data. 

This  paper  presents  an  initial  exploration  into  the  missing  forestry  data 
problem  given  the  autocorrelated  nature  of  most  spatial  ecological  data.  Space 
restrictions  preclude  a  complete  discussion  of  this  topic. 

SPATIAL  STRUCTURE 

While  the  problem  outlined  above  may  seem  to  be  a  simple  statistical 
inference  problem,  it  is  not.  The  distribution  of  natural  abiotic,  and  induced  biotic 
features,  are  typically  neither  uniform  nor  random.  Rather,  they  tend  to  be  either 
aggregated  in  patches  or  form  gradients  or  other  types  of  spatial  structures 
(Legendre  1993). 

Why  is  representation  of  spatial  pattern  important  in  forest  and  ecological 
models?  The  spatial  or  temporal  structure  or  pattern  of  ecosystems  is  a  crucial 
element  in  most  ecological  theories.  Any  forest  or  ecological  model  based  on  such 
theories  must  take  spatial  pattern  and  the  nature  of  spatial  ecological  data  into 
account.  Specific  characteristics  of  spatial  data  can  be  identified  which  cause 
problems  in  traditional  statistical  techniques.  There  are  two  main  classes  of  spatial 
effects:  spatial  dependence  and  spatial  heterogeneity  (Anselin  and  Getis  1992, 
Cressie  and  ver  Hoef  1993).  Only  spatial  dependence  will  be  considered  in  this 
paper. 

Spatial  Autocorrelation 

What  is  significant  about  the  patterns  found  in  spatial  ecological  data? 
According  to  Tobler's  first  law  of  geography,  everything  is  related  to  everything 
else,  but  near  things  are  more  related  than  distant  things.  This  simply  means  that 
the  relations  within  and  between  variables  sampled  at  spatially  proximal  locations 
will  be  stronger  than  those  sampled  at  further  distances.  Ecological  data  are 
spatially  dependent  by  nature.  Spatial  dependence  is  often  referred  to  as  spatial 
autocorrelation. 

A  major  consequence  of  spatial  autocorrelation  is  that  traditional  statistical 
inference  techniques  are  not  as  efficient  as  an  independent  sample  of  the  same 
size.  That  is,  spatial  dependence  leads  to  a  loss  of  information,  information  which 
is  crucial  for  predicting  missing  values.  Positive  spatial  autocorrelation  may  cause 
the  variance  to  be  underestimated.  Negative  autocorrelation  may  inflate  estimates 
of  variance.  In  either  case,  predictive  equations  can  become  biased  and  the  results 


594 


of  prediction  unreliable. 

Clearly,  some  new  methods  or  enhancements  to  existing  statistical  methods 
of  spatial  inference  are  needed  to  address  the  case  of  spatially  autocorrelated  data 
(Antle  and  Klinkenberg  1995). 

MODELING  SPATIAL  AUTOCORRELATION 

As  discussed  above,  many  of  the  basic  statistical  methods  used  in  ecological 
studies  are  impaired  by  autocorrelated  data.  Most,  if  not  all,  environmental  data 
falls  into  this  category  (Legendre  and  Fortin  1989).  Predict  ion  methods  which  can 
be  used  to  fill  in  missing  forestry  data  must  account  for  spatial  autocorrelation  or 
be  resistant  to  its  effects.  Before  these  techniques  are  applied,  the  nature  of  the 
spatial  patterns  in  the  data  should  be  explored.  Various  tests  for  autocorrelation 
and  trends  can  be  used.  Once  we  know  that  the  data  is  correlated  and  the  nature 
of  the  relation  (i.e.,  either  positive  or  negative),  techniques  which  take  this  pattern 
into  account  can  be  applied. 

Testing  for  Autocorrelation 

Since  the  presence  of  autocorrelation  may  distort  the  results  of  conventional 
statistical  interpolation  techniques,  autocorrelation  should  be  tested  for  prior  to 
selecting  any  particular  model  (i.e..  during  the  exploratory  phase  of  data  analysis) 
in  order  to  ensure  accurate  predictions.  There  are  various  methods  which  test  for 
the  presence  of  spatial  autocorrelation  in  univariate  data  and  for  correlation  in 
multivariate  data  sets.  These  are  well  documented  and  will  not  be  described  here. 
Univariate  methods  include  the  joint  count  test  for  nominal  data.  Geary's  c  or 
Moran's  /  test  for  ordinal  and  interval  data,  and  spectral  analysis  (see  Haining 
1990  for  a  detailed  description).  Multivariate  methods  include  the  Spearman  rank 
correlation  coefficient  for  ordinal  data,  the  Pearson  product  correlation  coefficient 
for  interval  data,  the  Mantel  test  and  correlogram  (see  Haining  1990  or  Goodchild 
1986  for  details). 

If  large  scale  spatial  dependency  exists,  it  may  be  removed  by  regression  or 
model-fitting  (eg.,  trend  surface  analysis).  Traditional  statistical  techniques  can 
then  be  used  without  violating  their  assumptions.  However,  caution  should  be 
taken  in  order  to  avoid  removing  the  determinants  of  the  underlying  processes 
along  with  the  spatial  structure.  An  alternative  to  removing  spatial  dependence  is 
to  modify  statistical  approaches  to  take  spatial  autocorrelation  into  account.  At  the 
very  least,  techniques  which  are  resistant  to  spatial  autocorrelational  effects  should 
be  used.  Techniques  which  utilize  autocorrelational  information  (to  enhance 
prediction  accuracy)  are  preferable.  First  how  do  we  represent  spatial 
autocorrelation0 


595 


Representing  Autocorrelation 


Autocorrelation  can  be  represented  through  structure  functions,  which  help 
quantify  spatial  dependency  and  partition  it  along  distance  intervals. 

Geographic  coordinates 

The  spatial  structure  of  variables  can  be  expressed  as  a  linear  combination  of 
geographic  coordinates  of  sample  sites  (Legendre  1993).  For  example,  a  high  order 
polynomial  can  be  built  up  using  the  x  and  y  coordinates  of  sites  (as  is  done  in 
trend  surface  analysis).  This  information  can  then  be  incorporated  into  traditional 
inference  models,  such  as  regression  models,  using  partial  regression  analysis 
techniques.  The  result  of  this  analysis  is  the  separation  of  the  variation  of  the 
target  variable  into  four  components:  the  variance  resulting  from  non-spatial 
environmental  factors,  spatial  structure,  the  interaction  of  environmental  and  spatial 
factors,  and  the  unexplained  variation. 

Geographic  distances 

Another  approach  is  to  use  location  as  one  of  a  set  of  predictor  variables  in 
a  statistical  model  (Legendre  1993).  In  this  case,  the  spatial  structure  of  the  data 
is  represented  by  a  matrix  of  geographic  distances  between  site  locations  (also 
called  a  proximity  matrix).  For  example,  the  euclidean  distance  can  be  computed 
for  all  pairs  of  sites  based  on  their  geographic  coordinates.  This  information  is 
then  summarized  in  a  spatial  distance  matrix.  If  the  remaining  environmental 
variables  can  be  represented  in  the  form  of  distance  matrices,  then  these  distance 
matrices  can  be  compared  using  some  form  of  correlation  analysis  (e.g.,  the 
Mantel  test,  see  Legendre  and  Fortin  1989). 

Covariation 

One  of  the  most  common  structure  functions  is  the  variogram  (and  related 
covariance  and  correlogram).  The  variogram  summarizes  the  spatial  continuity  for 
all  possible  pairings  of  data,  for  all  significant  lag  distances,  by  modeling  the 
average  degree  of  similarity  between  values  as  a  function  of  their  separation 
distance.  Variograms  can  be  computed  either  as  averages  over  all  directions  or 
specific  to  a  particular  direction.  If  there  is  a  trend  in  the  data  (i.e.,  local  means 
and  local  variances  change  as  a  function  of  location  within  the  sampling  space), 
the  variogram  will  include  both  lag-to-lag  variability  and  the  trend  variability 
(Rossi  et  al.  1992).  In  the  case  of  stationary  data,  the  variogram  can  be 
experimentally  estimated  from  observed  values.  Various  statistical  models  can  then 
be  fitted  to  the  variogram,  including  exponential,  spherical  and  linear. 

Prediction  Utilizing  Autocorrelation 

Since  ecological  data  are  often  correlated,  techniques  which  can  take 


596 


advantage  of  the  information  contained  in  these  relationships  may  be  able  to  best 
predict  missing  values  while  retaining  the  variability  and  covariation  structure  of 
forest  attribute  data. 

Regression  models 

Traditional  linear  regression  can  be  extended  to  situations  where  spatial 
autocorrelation  exists  by  the  addition  of  lagged  variables.  Griffith  (1987)  outlines 
how  spatial  autocorrelation  can  be  incorporated  into  traditional  regression  models. 
Promising  results  have  been  obtained  from  the  introduction  of  an  additional 
parameter  into  conventional  statistical  models  that  accounts  for  the  latent  spatial 
autocorrelative  structure  of  the  data.  This  technique  has  also  been  extended  to 
multivariate  models  (Haining  1990).  This  approach  is  one  step  toward  the 
integration  of  spatial  dependence  (in  the  form  of  geographic  distance  information) 
into  predictive  models. 

As  mentioned  above,  in  partial  regression  analysis,  the  spatial  component  can 
be  "partialled  out"  by  regressing  the  spatial  factor  variables  onto  each  explanatory 
variable  until  only  the  regression  residuals  remain  (Legendre  1993).  The  residuals 
are  then  used  to  model  the  variable(s)  which  have  missing  values.  This  approach 
"explains"  much  of  the  variation  that  is  left  unaccounted  for  in  simple  regression 
models. 

Kriging 

Kriging  is  a  well  known  geostatistical  spatial  interpolation  technique  which 
explicitly  uses  autocorrelational  information  in  its  predictions  (Rossi  et  al.  1994). 
Since  it  uses  a  local  estimator  and  the  autocorrelation  structure  of  the  data,  kriging 
is  quite  unlike  traditional  methods,  such  as  trend  surface  analysis,  which  ignore 
autocorrelational  effects  and  use  all  of  the  data  to  estimate  the  unknown  point's 
value.  In  addition,  kriging  provides  variance  information  which  is  crucial  in 
determining  the  accuracy  of  predicted  values. 

Kriging  can  be  applied  to  both  nominal  and  continuous  data.  In  general, 
kriging  predicts  missing  values  at  a  specific  site  by  taking  a  weighted  linear 
average  of  available  samples  (like  regression).  It  is  similar  to  multiple  regression 
with  a  few  important  differences.  First,  kriging  can  produce  predicted  values  that 
are  either  larger  or  smaller  than  any  of  the  sample  values.  Second,  kriging  utilizes 
both  distance  and  geometry  among  samples,  whereas  traditional  methods  use  only 
distance.  Third,  kriging  attempts  to  minimize  the  variance  of  the  expected  error  by 
inferring  the  variance  from  an  empirical  model  of  degree  of  spatial  dependence 
with  distance  and  direction  (i.e.,  the  variogram). 

Most  similar  neighbor  analysis 

Most  similar  neighbor  (MSN)  analysis  is  a  canonically  based  technique  which 
can  be  used  to  fill  in  interval  or  continuous  data  values  by  implicitly  utilizing  the 
spatial  structural  information  of  the  data.  Instead  of  estimating  missing  values  one- 


597 


by-one  (for  a  particular  site),  the  technique  "chooses"  a  most  similar  site  from  a 
set  of  sites  to  act  as  a  stand-in.  The  surrogate  site  is  chosen  on  the  basis  of  a 
similarity  function  (derived  using  canonical  correlation  analysis). 

This  approach  works  well  when  the  variable  to  be  predicted  and  the  predictor 
variables  have  strong  functional  relationships  to  each  other  (Moeur  1995)  (as  is 
the  case  for  spatially  autocorrelated  data).  Because  surrogate  sites  are  chosen  from 
among  actual  sites,  impossible  predictions  for  missing  values  cannot  occur. 
However,  this  means  that  the  samples  sites  must  adequately  span  the  full  range  of 
sites.  While  Moeur's  procedure  does  not  explicitly  model  spatial  correlation,  it 
attempts  to  maintain  the  covariance  structure  of  the  multivariate  response 
attributes. 

DISCUSSION 

None  of  the  methods  described  here  for  filling  in  missing  forestry  data  can 
be  used  without  careful  consideration.  They  all  make  assumptions  about  the  nature 
of  data  and  the  phenomena  under  study.  For  specific  cases,  these  assumptions  may 
or  may  not  hold.  The  choice  of  techniques  should  consider  how  the  information 
contained  in  the  observed  data  can  be  used,  the  accuracy  of  predictions,  the 
amount  of  computational  effort  required,  and  how  flexible  they  are  in  allowing 
both  theoretical  and  empirical  information  to  be  utilized  (Haining  1990). 

Regression  can  be  extended  by  incorporating  a  spatial  autocorrelation 
parameter  and  by  using  different  forms  of  the  model  with  spatially  lagged 
variables.  A  fair  amount  of  trial  and  error  is  required  to  determine  the  best  fitting 
model  for  a  particular  data  set.  For  a  particular  study,  the  different  regression 
models  should  be  compared  not  only  in  terms  of  included  and  excluded  variables 
but  also  in  terms  of  different  forms  of  the  model.  Interpretive  knowledge  can  be 
utilized  in  the  choice  of  model. 

Regression  assumes  normality  and  stationarity,  which  may  or  may  not  hold. 
While  regression  is  useful  for  estimating  population  means  and  totals,  it  does  not 
always  preserve  the  underlying  relations  among  inventory  attributes.  For  example, 
if  one  or  more  variables  with  missing  values  is  difficult  to  predict,  the  correlation 
structure  may  be  distorted.  Regression  may  produce  unreliable  results  for  cases  of 
high  unique  variances. 

While  traditional  methods  of  interpolation  (e.g.,  inverse  distance  weighting, 
local  sample  means)  are  less  accurate,  they  are  also  less  computationally  intensive 
than  kriging  and  do  not  require  the  subjective  estimation  of  the  variogram.  Kriging 
assumes  that  the  variogram  is  accurate  over  the  entire  study  area.  This  is  known 
as  the  stationarity  hypothesis  and  in  cases  where  it  does  not  hold,  the  results  from 
kriging  will  be  less  accurate.  In  addition,  kriging  assumes  that  the  variable  being 
predicted  is  multivariate  normal.  Raw  data  are  rarely  univariate  normal  (Rossi  et 
al  1994). 

While  MSN  analysis  is  a  robust  approach  that  implicitly  captures  spatial 


598 


structure,  it  sensitive  to  sample  size  in  that  smaller  samples  sizes  reduce  the 
variance,  which  in  turn  affects  it's  predictive  function.  Results  from  MSN  analysis 
are  unpredictable  if  a  strong  non-linear  relation  exists  among  the  variables  of  one 
or  both  sets.  Canonical  analysis  (upon  which  MSN  is  based)  is  adequate  for  the 
case  where  the  joint  distribution  of  variables  is  linear  and  continuous.  So,  in  cases 
where  data  are  either  heterogeneous  or  exhibit  non-linear  relations  between 
variables,  MSN  will  fall  short.  As  well,  the  linear  orthogonal  nature  of  solutions 
in  canonical  analysis  may  be  unrealistic  for  ecological  data  where  there  is  often 
high  unique  variances. 

An  issue  which  is  important  but  has  not  been  addressed  in  this  paper  is  that 
of  the  scale  of  variation.  Spatial  variation  may  be  conceptualized  as  being 
composed  of  several  (possibly  independent)  components  at  several  different  scales 
(Haining  1990).  In  testing  for,  representing  and  modeling  spatial  autocorrelation, 
it  is  important  to  identify  and  account  for  the  contribution  from  each  of  the 
different  scales  to  the  total  variation. 

CONCLUSIONS 

If  we  rely  on  prediction  models  that  assume  that  biological  features  are 
distributed  uniformly  or  randomly  in  space,  then  chances  of  obtaining  accurate 
predictions  of  missing  values  are  small,  since  spatial  autocorrelation  characterizes 
most  ecological  data.  It  is  often  difficult  to  determine  if  assumptions  have  been 
broken,  making  it  hard  to  compensate  for  such  violations  or  to  know  how  reliable 
the  newly  completed  data  set  is.  Consequently  spatial  autocorrelation  must  be 
tested  for  and  techniques  which  either  explicitly  or  implicitly  consider  spatial 
autocorrelation  must  be  used  when  it  is  present. 

While  the  different  techniques  examined  above  have  various  advantages, 
disadvantages  and  restrictions,  it  is  impossible  to  produce  a  truly  meaningful 
comparison  unless  each  method  is  applied  to  a  unique  set  of  variables.  To  date 
there  have  been  few  studies  which  have  formally  compared  the  various  spatial 
inference  techniques  which  can  be  used  to  fill  in  missing  data  values.  Given  the 
increased  importance  such  methods  will  have  in  the  future,  conducting  such  a 
comparative  study  is  certainly  warranted. 

REFERENCES 

Antle,  A.  N.  and  Klinkenberg,  B.  1995.  Statistically-based  data  generation 

techniques:  an  emerging  trend.  In:  Heit,  M.,  Parker,  H.  D.  and  Shortreid,  A. 

(eds.)  GIS  Applications  in  Natural  Resources  2,  GIS  World,  Inc. 
Anselin,  L.  &  Getis,  A.  1992.  Spatial  statistical  analysis  and  geographic 

information  systems.  The  Annals  of  Regional  Science,  Springer- Verlag,  Vol. 

26,  pp.  19  -33. 

Chrisman,  N.R.  1991.  The  error  component  in  spatial  data.  In:  MacGuire  D.J., 


599 


Goodchild,  M.F.  &  Rhind,  D.W.  (eds.)  Geographical  Information  Systems: 
Principles  and  Applications,  Longman  Scientific  &  Technical,  Vol.  1,  NY. 
Cressie,  N.  &  ver  Hoef,  J.M.  1993.  Spatial  statistical  analysis  of  environmental 
and  ecological  data.  In:  Goodchild,  M.F.,  Parks,  B.O.  &  Steyaert,  L.T.  (eds.), 
Environmental  modeling  with  GIS,  pp.  404  -  421.  Oxford  University  Press 
Inc.,  Oxford. 

Goodchild,  M.  F.  1986.  Spatial  autocorrelation.  CATMOG  47,  Geo  Books, 
Norwich. 

Goodchild,  M.  &  Gopal,  S.  1989.  Accuracy  of  spatial  databases.  Taylor  & 

Francis,  Ltd.,  London. 
Griffith,  D.A.  1987.  Spatial  autocorrelation  :  A  primer.  Monograph  of  the 

American  Association  of  Geographers,  Washington. 
Guptill,  S.C.  &  Morrison,  J.L.  1995.  Looking  Ahead.  Morrison,  J.L.  (eds.) 

Elements  of  Spatial  Data  Quality,  Elsevier  Science,  Oxford,  UK. 
Haining,  R.   1990.  Spatial  data  analysis  in  the  social  and  environmental 

sciences.  Cambridge  University  Press,  Cambridge. 
Legendre,  P.  &  Fortin,  M.J.  1989.  Spatial  pattern  and  ecological  analysis. 

Vegetatio,  Vol.  80,  pp.  107  -  138. 
Legendre,  P.  1993.  Spatial  autocorrelation:  trouble  or  new  paradigm?  Ecology, 

Vol.  74,  No.  6,  pp.  1659  -1673. 
Moeur,  M.  and  Stage,  A.  R.  1995.  Most  similar  neighbor:  an  improved 

sampling  inference  procedure  for  natural  resource  planning.  Forest  Science, 

Vol.  41,  No.  2,  pp.  337-359. 
Rossi,  R.E.,  Mulla,  D.J.,  Journel,  A.G.  &  Franz,  E.H.  1992.  Geostatistical  tools 

for  modeling  and  interpreting  ecological  spatial  dependence.  Ecological 

Monographs,  Vol.  62,  No.  2,  pp.  277  -  314. 
Rossi,   R.E.,   Dungan,  J.L.   &  Beck,  L.R.    1994.   Kriging  the  shadows: 

geostatistical  interpolation  for  remote  sensing,  Remote  Sensing  Environment, 

Vol.  49,  pp.  32  -  40. 

BIOGRAPHICAL  SKETCH 

Alissa  N.  Antle  is  a  doctoral  student  in  the  Department  of  Geography  at  the 
University  of  British  Columbia.  She  graduated  from  the  University  of  Waterloo 
with  a  B.A.Sc.  in  Systems  Design  Engineering  and  a  B.A.  in  Liberal  Arts.  Alissa's 
research  interests  include  visualization  and  exploration  of  spatial  data  quality  and 
using  GIS  for  decision  making  in  ecosystem  management  environments. 

Peter  L.  Marshall  is  an  Associate  Professor  in  the  Department  of  Forest 
Resources  Management  at  the  University  of  British  Columbia.  He  received  B. Sc. F. 
and  M.Sc.F.  degrees  in  forestry  from  the  University  of  Toronto  and  a  Ph.D.  in 
forest  management  from  the  University  of  British  Columbia.  His  research  interests 
include  sampling  designs  for  collecting  resource  data  and  approaches  to 
quantifying  and  predicting  stand  dynamics. 


600 


Optimizing  Sampling  Schemes  for  Mapping 
and  Dredging  Polluted  Sediment  Layers 


L.  Hazelhoff1  and  F.  Hoefsloot1 

Abstract.-An  optimal  sample  scheme  of  measurements  of 
sediment  depth  was  constructed  for  mapping  and  dredging  a 
polluted  sediment  layer,  taking  into  account  economic  constraints. 
The  layer  in  question  was  deposited  in  a  lake  near  Kampen  in  the 
Netherlands  and  consists  of  sediment  discharged  by  the  Rhine 
during  the  period  1932  -1980.  A  stochastic  approach  was  chosen  to 
estimate  the  depth  of  dredging  that  would  result  in  a  given  amount 
of  polluted  sediment  remaining  in  the  lake  after  dredging  has  been 
carried  out.  The  dredging  result,  i.e.  the  amount  of  sediment  that  is 
not  removed,  depends  on  the  probability  distribution  of  the 
estimated  depth  and  the  difference  between  the  mean  of  this 
distribution  and  a  chosen  dredging  depth.  Two  independent  sources 
of  error  determine  this  probability  distribution:  i)  the  interpolation 
error  and  ii)  the  precision  of  the  determination  of  the  depth  of 
dredging.  The  two  sources  of  error  are  independent.  Spatial 
interpolation  was  done  using  ordinary  point  kriging.  The  kriging 
error  depends  on  the  sampling  scheme  and  the  spatial  variability  of 
the  depth  of  the  bottom  of  the  polluted  sediment  layer.  Spatial 
correlation  at  three  different  scales  was  modelled  with  a  nested 
variogram.  The  sampling  distance  of  the  sediment  depth 
measurements  was  evaluated  as  a  function  of  dredging  result  and 
dredging  depth  precision.  This  information  can  be  used  to  derive  a 
relation  between  sampling  costs  and  dredging  result.  The  amount 
of  unnecessarily  dredged  clean  sediment  was  estimated  and 
incorporated  in  the  procedure  for  designing  an  optimal  sample 
scheme. 


INTRODUCTION 

In  the  period  1932  -  1980  polluted  Rhine  sediments  were  deposited  in  lake 
Ketelmeer.  located  near  Kampen  in  the  Netherlands.  This  polluted  sediment  now 
has  to  be  removed  from  an  area  of  about  30  kirT.  A  pilot  study  was  carried  out  in 
two  small  areas  to  get  an  impression  of  the  spatial  variability  of  the  depth  of  the 
under  side  of  this  layer.  The  sampling  was  proceeded  in  stages,  with  information 
from  the  first  area  being  used  for  sampling  the  second.  Within  the  pilot  areas 
different  sample  stages  were  also  applied. 

'Physical  Geographers,  University  of  Utrecht,  Tin-  Netherlands. 


601 


Dredging  companies  have  recently  used  triangulation  (TIN)  to  construct  a  digital 
model  of  the  level  of  dredging,  and  in  some  cases  the  spatial  variability  was 
estimated  using  spectral  waves  or  ARIMA  models.  The  results  of  these  studies 
were  used  to  increase  the  depth  of  dredging  below  the  TIN  surface  to  ensure  that 
downward  fluctuations  in  the  level  of  the  polluted  sediment  will  be  removed.  The 
sampling  distance  was  arbitrarily  determined.  The  main  objective  of  this  study  is 
to  estimate  the  extra  dredging  depth  with  reference  to  an  interpolated  DEM  using 
Kriging  and  the  Kriging  variance.  Using  the  Kriging  variances,  a  relation  can 
established  between  extra  dredging  depth,  sample  distance  and  the  amount  of 
polluted  sediment  that  is  not  removed  by  dredging.  A  second  objective  is  to 
extrapolate  the  spatial  variability  of  the  polluted  layer  in  two  pilot  areas  to  the 
whole  lake  so  that  economic  assessments  can  be  made  of  the  sample  scheme 
needed  for  the  lake  for  a  given  fraction  of  polluted  sediment  that  is  not  removed. 
A  problem  in  this  area  is  non-stationary  spatial  variation  caused  by  human  activity 
in  the  past  and  today.  Digging  of  clay  and  sand  has  resulted  in  pits  have  mostly 
been  filled  with  polluted  sediment,  and  along  the  shipping  channels  vessels  cause 
resuspension  and  redeposition  of  the  sediment  along  the  channels.  Both 
disturbances  must  be  identified  as  outliers  in  the  data,  so  careful  exploratory 
analysis  of  the  spatial  data  is  necessary.  First  the  spatial  variability  of  the  two 
samples  areas  is  explored,  then  the  method  of  determining  the  economically 
optimal  dredging  in  relation  with  environmental  constraints  is  explained  and 
applied  on  the  results  of  the  second  sample  area. 

The  study  was  funded  by  the  Ministry  of  Public  Works  (RWS). 

Area  one 

sample  scheme 

This  area  is  800x800  m  and  was  chosen  by  RWS  as  a  test  location  for 
redevelopment  using  new  methods  of  dredging.  Based  on  information  from  a 
survey  on  a  slightly  different  sediment  layer  in  the  same  area  the  following  sample 
scheme  was  designed.  Four  subareas  of  100x100  m  were  randomly  chosen 
containing  sample  grids  of  20  meter  (figure  1).  This  approach  can  be  justified  by 
the  knowledge  of  the  sedimentation  process;  it  is  assumed  that  the  spatial 
variability  at  this  scale  in  the  four  selected  blocks  will  be  representative  for  the 
whole  lake.  To  investigate  whether  the  subareas  are  influenced  by  variation  at  a 
larger  scale,  a  100  meter  grid  was  sampled  in  the  whole  area.  A  third  level  of  scale 
was  investigated  at  sample  distances  ranging  from  0.5  to  5  m,  realized  by  three 
groups  of  two  short  transects  designed  with  the  shape  of  a  cross.  The  main 
objective  of  this  short  distance  sampling  was  to  test  measurement  accuracy;  at 
these  distances  spatial  variation  was  assumed  lower  than  expected  measurement 
error,  RWS  planned  this  lowest  level  of  sampling.  The  information  was  also  used 
to  make  a  better  estimate  of  the  nugget. 


602 


513700 


£ 

v  * 


513000 


181000        181200       181400       181600       181800  182000 

X(m) 


-500  -475  -450  -425   -400  -375  -350  -325  -300 

Depth  (cm) 

Figure  1.   Sample  design  and  interpolated  surface  of  area  one. 


Gamma  (h)  :  4  +  124*Gauss  63  (h)  +  1.2*Pow  1  (h) 


Omnidirectional 


figuur  2.   Semivariogram  of  the  under  side  of  the  polluted  layer  in  area  one. 


603 


results 

In  this  area  variation  over  distances  of  100-500  m  is  clearly  present  with  two  of 
the  subareas  being  located  in  zones  with  a  trend.  Performing  an  analysis  on 
residuals  reveals  the  variability  at  the  20-100  m  scale,  but  spatial  outliers  were 
found  in  these  subareas,  using  exploratory  data  analysis  (EDA)  software.  The 
EDA  software  visualizes  spatial  variability  by  connecting  measurement  locations 
to  lines  which  can  be  selected  from  a  graph  displaying  the  variogram  cloud 
(Hazelhoff  &  Gunnink,  1992).  These  spatial  outliers  were  identified  as  being 
caused  by  human  interferences,  and  so  they  could  not  be  used  for  investigating 
spatial  variation  at  the  20-100  m  scale.  The  variogram  in  figure  2  is  based  on  2  x 
25  samples  in  this  range.  A  nested  variogram  has  been  constructed;  the 
experimental  variograms  for  three  spatial  scales  were  calculated  separately  and 
then  combined  for  fitting.  It  shows  a  low  nugget  caused  by  the  limited  spatial 
variation  at  the  level  of  the  crosses  (0.5  to  5  m).  A  steep  slope  between  0  and  60m 
is  clearly  present. 

Area  two 

sample  scheme 

In  this  area  measuring  400x500  meter  (figure  3)  a  dredging  test  was  planned,  so  a 
DEM  had  to  be  constructed.  A  grid  mesh  of  50m  was  planned,  based  on  the 
variogram  of  area  one  with  the  steep  increase  in  the  range  of  0  -  60  m,  so  much 
lower  kriging  variances  will  occur  when  a  50  m  grid  mesh  is  used  instead  of  100 
m.  Due  to  practical  constraints  the  realized  sample  scheme  is  somewhat  different. 
In  this  area  one  subarea  of  100x100  m  was  sampled  at  distances  of  20  m  located 
on  a  grid  to  compare  the  results  with  area  one.  Furthermore  two  crosses  (with  the 
same  configuration  as  in  area  one)  were  added  with  sampling  distances  ranging 
from  0.5  to  5  m.  A  second  sampling  has  been  carried  out  for  validation  of  the 
DEM.  These  samples  also  are  plotted  in  figure  3,  many  of  them  are  located  close 
to  the  samples  taken  in  the  first  stage. 


604 


179900       180000       180100       180200  180300 

X(m) 


-360  -350  -340  -330 

Depth  (cm) 

Figure  3.  Sample  scheme  and  interpolated  surface  of  area  two. 


Gamma  (h) :  9  +  10*Sph  30  (h)  +  0.01 3*Pow  1.53  (h) 


605 


results 


After  using  EDA  to  identify  human  induced  spatial  outliers,  the  spatial  variability 
appeared  to  be  much  lower  than  in  areas  one.  The  second  stage  of  sampling  for 
validation  confirms  the  originally  variogram,  except  the  nugget  variance  appears 
now  too  low.  Pairs  of  samples  close  together  (<10  meter)  and  are  not  in  the 
crosses  show  much  higher  semivariances  then  samples  within  the  crosses.  Two 
possible  reasons  can  be  given:  1)  the  crosses  are  located  in  areas  with  low  spatial 
variability  at  short  distances,  and  2)  measurement  errors  are  lower  within  the 
crosses  then  outside  the  crosses.  The  variogram  based  on  all  samples  is  showed  in 
figure  4.  The  steepest  part  ranges  from  10  to  30  m,  which  is  the  important  interval 
for  kriging  the  50  meter  grid.  The  right  part  of  the  variogram  fitted  with  a  power 
model  represents  some  linear  trend  in  this  area.  For  coarser  grids  this  trend  will  be 
important. 


The  DEM  of  the  lower  side  of  the  polluted  layer  in  area  two  was  made  using 
ordinary  point  kriging  (OPK).  During  the  dredging  tests  the  dredging  depth  was 
set  at  the  level  of  the  interpolated  DEM  which  means  that  polluted  sediment  is 
still  present  (residual  sediment)  in  half  of  the  area,  assuming  a  normal  distribution 
of  the  deviations.  The  volume  of  the  residual  sediment  can  be  calculated  if  the 
distribution  of  the  deviations  is  known.  In  this  study  the  kriging  variance  (normal 
distributed)  and  the  error  of  estimating  the  dredging  depth  are  used  to  describe  the 
deviations  from  the  DEM.  These  two  sources  of  error  are  supposed  to  be 
independent,  simply  adding  the  variances  gives  the  total  error.  Increasing  the 
dredging  depth  relative  to  the  DEM  results  in  smaller  volumes  of  residual 
sediment.  The  amount  of  residual  sediment  is  calculated  using  the  tail  of  the 
normal  probability  function.  For  example,  if  a  =  1  cm  then  increasing  the 
dredging  depth  by  1  cm  results  in  a  coverage  of  16%  of  the  area  with  polluted 
sediment.  The  residual  volume  V  per  unit  of  area  can  be  calculated  using  (1). 


z  is  the  depth  relative  to  the  DEM  where  the  center  of  the  distribution  lies  and  p(z) 
is  the  probability  density  of  the  normal  distribution.  Integration  starts  at  dredging 
depth  Zd>  This  integral  is  discretised  to  calculate  V.  The  quantity  V  depends  on  a, 
the  kriging  standard  deviation  (KSD)  and  Zd-  KSD  depends  on  the  variogram  and 
the  sample  scheme.  The  dredging  result  is  defined  as  the  result  of  dividing  the 
residual  volume  by  the  total  area. 


Dredging  depth  &  residual  sediment 


(1) 


Zd 


606 


For  area  two  the  extra  dredging  depth  (EDD)  was  calculated  for  a  range  of 
dredging  results,  and  is  repeated  for  different  values  of  the  dredging  depth  error. 
During  the  calculation  each  grid  cell  gets  a  different  value  for  V  because  KSD 
varies  over  the  area.  A  special  computer  program  was  written  to  work  out  this 
iterative  procedure.  Another  interesting  quantity  is  the  volume  of  clean  sediment 
dredged  as  function  of  EDD.  Dredging  at  DEM  depth  means  that  dredged  clean 
sediment  (DCS)  is  equal  to  V.  The  volume  of  dredged  polluted  sediment  is  equal 
to  V(Z(1=Zdem)  -  V(Zd)-  The  volume  of  clean  sediment  is  calculated  from  formula  2. 
EDD  is  here  expressed  as  a  volume  with  an  area  of  1 . 

DCS  =  EDD  -  ( V(zd=ZDEM)  -  V(Zd) )  +  V(zd=ZDEM)  =  EDD  +  V(Zd).  (2) 

It  is  obvious  from  (2)  that  increasing  EDD  at  low  values  of  V(zd)  yields  small 
improvement  in  the  dredging  result  relative  to  the  increase  of  DCS. 

DREDGING  RESULT  &  SAMPLE  SCHEME 

For  economic  reasons  the  most  interesting  question  is  how  does  the  EDD  depend 
on  the  sample  distance  (regular  grid)  given  a  certain  dredging  result  and  a  given 
the  spatial  variability  described  with  a  variogram?  For  the  variogram  of  area  two 
this  relationship  was  calculated  and  plotted  in  figure  5.  For  three  sample  distances 
(100,  50  and  25  m)  the  relation  between  dredging  standard  deviation  DSTD  and 
EDD  is  plotted.  The  dredging  result  is  set  at  1  cm,  which  is  a  practical  value. 
DSTD  is  an  important  error  source  because  the  values  are  comparable  with  the 
KSTD  based  on  the  three  sample  distances.  Improvement  of  estimating  dredging 
depth  results  in  a  significantly  lower  EDD.  At  the  moment  some  dredging 
companies  claim  values  for  DSTD  near  2  cm.  It  is  up  to  RWS  to  balance 
increasing  cost  due  to  extra  dredging  depth  against  the  increasing  cost  of  denser 
sampling. 


607 


6 


u 

Q 
Q 


10  20  30  40 

DREDGING  STANDARD  DEVIATION  (MM) 


50 


25  M 
50  M 
100  M 


Figure  5.  Relation  between  dredging  standard  deviation  and  extra  dredging  depth 

with  a  given  dredging  result  of  1  cm. 

Extrapolation  from  pilot  areas  to  the  whole  lake 

The  relations  between  sample  density  and  EDD  were  derived  from  the  variogram 
of  area  two.  The  variogram  in  area  one  differs  considerable  from  this  variogram; 
in  area  one  kriging  variances  are  larger  for  any  given  sample  distance.  But  the 
spatial  variability  in  area  two  is  most  likely  to  be  the  most  representative.  Area 
one  is  located  near  the  border  of  the  lake  with  a  lot  of  human  disturbance.  If  it  is 
true  that  area  two  is  representative  for  the  lake,  a  50  m  grid  uses  information  from 
the  spherical  part  of  the  variogram.  The  recommendation  to  RWS  is  to  start  with  a 
100  m  grid  to  explore  the  spatial  variability  at  a  larger  scale  than  50  m,  as  is 
clearly  present  in  the  two  areas.  In  areas  with  such  a  large  spatial  variability  that  a 
large  EDD  is  needed  to  match  the  desired  dredging  result,  the  50  m  grid  is 
advisable.  Another  important  fact  is  the  presence  of  human  disturbances  which 
need  dense  sampling  to  be  tracked.  There  is  a  risk  that  disturbances  of  at  most  1 
ha  will  be  missed  using  a  100  m  grid,  which  can  be  important  because  the 
thickness  of  the  polluted  sediment  in  the  areas  may  be  more  than  doubled. 


608 


CONCLUSIONS 


Spatial  correlation  at  two  distinctive  scales  separated  at  about  40  meter  is  present 
in  the  lower  boundary  of  the  polluted  sediment. 

Ordinary  point  kriging  can  be  used  together  with  an  error  specified  for  estimating 
dredging  depth,  to  optimize  sample  schemes  given  a  dredging  result.  This 
information  cannot  be  derived  from  TIN  models  and  associated  approximators  of 
variability  like  ARIMA  and  spectral  waves. 

For  the  spatial  variability  in  the  second  pilot  area,  dredge  depth  accuracy  is  an 
important  source  of  error. 

REFERENCES 

Hazelhoff,  L.  &  Gunnink,  J.L.  1992.  Linking  Tools  for  Exploratory  Analysis  for 
Spatial  Data.  In  Harts,  J.,  Ottens,  F.L.;  Conference  Proceedings  EGIS  1992. 

BIOGRAPHICAL  SKETCH 

Lodewijk  Hazelhoff  is  a  Physical  Geographer  specialized  in  environmental 
assessment  and  GIS.  He  graduated  from  the  University  of  Amsterdam  in  1980. 

Frans  Hoefsloot  is  a  Physical  Geographer  specialized  in  environmental 
assessment.  He  graduated  from  the  University  of  Utrecht  in  1992. 


609 


Forest  Cover  Monitoring  in  India  : 
The  Satellite  Experience 

J.B.  Lai' 

Abstract  :  In  certain  quantitative  aspects,  India  is  the  biggest  user  of 
satellite  imagery  in  the  monitoring  of  a  namral  resource.  The  Forest 
Survey  of  India  produces  a  biennial  forest  cover  map  of  the  entire 
countrv  based  on  interpretation  of  satellite  imagery.  The  interpretation 
which  was  in  the  beginning  manual,  is  now  a  combined  digital  and  visual 
effort.  The  area  covered  in  biennial  mapping  is  328  million  ha.  The 
average  annual  coverage  works  out  at  164  million  ha.  Apart  form  the 
resolution  of  the  sensor,  the  scale  of  mapping  significantly  influences  the 
accuracy  of  maps  The  Forest  Survey  of  India  has  so  far  produced  four 
series  of  maps.  The  first  series  was  produced  in  1987,  and  the  last  in 
1993.  The  1987  maps  were  on  1:  1  million  scale  based  on  visual 
interpretation  of  the  Landsat  MSS  imagery.  The  final  series  is  on 
1 :250.000  scale  based  on  interactive  interpretaion  of  Landsat  TM  and  IRS 
LISS  II  imagery.  From  the  first  to  the  fourth  of  the  series  of  the  maps 
there  is  a  bis  improvement  not  only  in  accuracy  but  also  in  the  utility  of 
the  maps. 

The  first  series  of  the  maps  was  useful  only  in  strategic  planning  at 
national  and  state  level,  while  the  fouth  series  is  useful  also  in  operational 
planning  at  district  (the  fundamental  administrative  unit  of  India)  level. 


INTRODUCTION 

A  few  years  from  entering  the  twenty-first  century,  India  still  lives  in  a  forest 
of  biomass  era.  It  still  uses  over  200  million  cubic  meter  of  fire-wood  for 
domestic  energy  and  still  lets  over  90  million  cattle  graze  in  forests.  One-sixth  of 
its  population,  recorded  as  scheduled  tribes  in  its  constitution,  still  meets  most  of 
its  living  needs  from  forests.  "(LAL  1989a)".  Forests  are  vital  to  India  not  only 
for  ecological  ,  but  also  for  scocio-economic  reasons.  Consequently,  rapid,  regular 
and  reliable  forest  cover  assessment  and  monitoring  is  an  indispensable  activity 
for  India. 


1  Former  Director.  Forest  Survey  of  India 


610 


Unfortunately,till  nineteen-eighties,  when  satellite  data  was  first  made  use  of  in 
forest  cover  assessment.  India,  at  no  point  of  time,  had  uptodate  information  about 
the  status  of  its  entire  forests.  All  it  knew  was  the  extent  of  area  legally  classified 
as  forests. 

Synoptic,  repetitive,  and  digital  as  it  is,  satellite  data  could  not  have  an 
alternative  to  itself  in  forest  cover  monitoring  in  a  large  and  diverse  country  like 
India,  where  hiotic  pressures  change  the  situation  of  a  natural  resource  in  months, 
let  alone  years  Even  on  1 :  50,000  scale  the  country  is  covered  by  93,000  aerial 
photographs,  while  only  in  200  sheets  of  satellite  imagery  from  Landsat  on 
1:250,000  scale  the  entire  country  could  be  viewed.  "(Lai  1989b)".  Further,  in 
India  which  contains  as  many  as  221  ecologically  stable  forest  types  which  vary 
from  tropical  wet  evergreen  forests  deep  in  the  south  to  alpine  vegetation  in  the 
north  in  the  Himalayas,  and  from  dry  thorn  forests  in  the  extreme  west  to  moist 
decidous  forests  on  the  eastern  border  "(Lai  1989b)"  ,  multispectral  and  multi- 
temporal  data  is  of  paramount  importance. 

Apart  from  technical  reasons,  satellite  data  also  saves  time  and  money  in 
obtaining  adequate  results  in  regard  to  forest  cover  assessment  and  monitoring. 

FOREST  COVER  MONITORING  IN  INDIA 

Using  the  satellite  data  the  Forest  Survey  of  India  (FST)  has  so  far  made  four 
biennial  efforts  in  assessing  forest  cover,  and  preparing  forest  vegetation  maps  of 
the  country  In  its  vegetation  mapping.  FSI  "(1988,1989,1993,1994)"  recognised 
only  four  categories  of  the  land  cover  : 

dense  forest  :  forest  of  40%  crown  density  or  more, 

open  forest  :  forests  of  less  than  40%  but  more  than  10%  crown  density 

scrub  forest  :  Tree-lands  with  less  than  10%  crown  density 

mangrove  forest  :  - 

non-forest  lands  :  - 

The  scale  of  mapping,  the  source  of  satellite  data,  the  sensor,  and  other  related 
information  in  regard  to  four  assessments  in  summarised  in  Table  1: 


611 


Tabic  1 


Assessment  Name  of    Name  of    The  period  to      Resolu-     Scale  of 


vear 

the 
Satellite 

the           which  imiwrv 

HIV                            T»  11  IV  11    IIIIllUVI  J 

sensor          roll*  tod 

tion 

mannina 

1087 

J->aIlU?Kll 

Mnlri-  1081-8^ 

7Qm 

1-1000  000 

Spectral 

Scanner(MSS) 

1989 

-do 

Thematic  1985-87 

30m 

1:250,000 

Mapper(TM) 

1991 

-do- 

-do-  1987-89 

-do 

-do 

1993 

Landsat 

TM  1989-91 

36m 

-do 

and 

and 

IRS 

LTSS  II 

Source  :  FSI  (198X  1989.1993,1994) 


The  interpretation  of  the  imagery  in  regard  to  first  three  assessments  was  purely 
manual,  but  the  last  was  a  combined  visual,  digital  effort.  The  results  obtained  in 
the  four  assessments  are  summarised  in  Table  2  : 


Tabic  2 

(in  million  hectare) 


Assessment        Forest  Scrub 
year      Dense    Open    Man-  Total 

grove 


Non-forest   uninte-  Total 

rpreted  land 

(clouds  area 
etc.) 


1987  36  14  27  66  0.40  64  20  7.68  255.74 

1989  37.85  25  74  0.42  64  01  6  61  257.77 

1991  38  50  24.99  0.42  61.91  5  97  256.93 

1993  38  56  25.03  0.42  64  01  5  90  258.82 


1.15  328.78 

0.39  328.78 

1.92  328.73 

Nil  328.73 


Source  :  FSI  (1988,1989  1993,1994) 


It  may,  however,  be  mentioned  that  the  comparison  of  1987  assessment  with 
other  three  assessments  is  not  totally  valid.  First,  in  the  1 987  assessment  the  forest 
cover  maps  were  prepared  on  1 : 1  million  scale  while  in  other  assessments  on 
1:250000  scale.  There  is  a  big  difference  in  the  cartographic  limits  of  the  two 
scales.  The  minimum  area  that  can  be  shown  on  1:1  million  scale  is  400  ha  while 
that  in  1:250.000  scale  it  is  25  ha.  In  other  words,  in  the  first  assessment,  the 
isolated  patches  of  forests  less  than  400  ha  in  extent  were  omitted  form  the 
counting.  In  the  same  manner  the  enclaves  of  other  land-uses  measuring  less  than 


612 


400  ha  in  forested  region  were  counted  as  forests.  Secondly,  the  imagery  used  for 
1987  assessment  was  from  Multi  Spectral  Scanner(MSS)  which  has  a  spatial 
resolution  of  79m  while  in  other  assessments  the  imagery  of  Thematic 
Mapper(TM)  which  has  a  spatial  resolution  of  30m  was  used. 

ACCURACY  OF  FOREST  ASSESSMENTS 

A  fact  which  needs  to  be  categorically  mentioned  is  that  in  classifying  forests 
in  'closed'  and  'open'  does  not  reveal  the  true  status  of  forests  in  regard  to 
composition  and  crown  density.  The  crown  density  of  a  forest  may  reduce  from 
80  to  50%,  the  forest  would  still  be  classed  as  'dense'.  And,  due  to  human 
intervention  a  forest  may  become  less  diverse  in  species  composition,  but  the 
satellite  imagery  would  say  little  about  it. 

It  should  also  be  appreciated  that  the  difference  in  radiance  value  of  a  land  unit 
at  two  points  of  time  mny  be  due  to  several  factors,  such  as  : 

*  change  in  vegetation 

*  differences  in  illumination/sun  angle 

*  differences  in  atmospheric  conditions 

*  differences  in  ground  moisture 

*  differences  in  sensor  caliberation 

Therefore,  radiance  change  on  account  of  change  in  crown  density  should  be 
relatively  large  for  imagery  to  give  authentic  information  about  land  cover. 

Though  forest  vegetation  maps  prepared  by  FST  in  various  assessments  were 
subjected  to  extensive  ground  truth  verification,  no  systematic  accuracy  assessment 
exercise  was  undertaken  by  the  organisation.  Lai  and  Prajapati(1989),  however, 
carried  out  a  reliable  exercise  in  Nagaland  state  to  ascertain  the  effect  of  scale  on 
the  accuracy  of  vegetation  mapping  from  satellite  data.  They  reported  that  over-all 
classification  accuracy  in  case  of  vegetation  mapping  on  1:1  million  scale  from 
MSS  data  was  only  54%,  while  on  1:250,000  scale  from  TM  data  it  was  79%. 

The  FSI  in  an  informal  communication  to  the  author  have  reported  that  digital 
interpretation  of  satellite  data  has  increased  the  classification  accuracy  to  90%. 
The  location  error  of  a  point  did  not  go  beyond  30  m. 

GRASSLAND  ASSESSMENT 

Apart  from  assessing  and  monitoring  forest  vegetation,  the  space-borne  data  has 
not  been  made  much  use  of  in  assessing  and  monitoring  other  land  uses.  A 
paramount  need  in  the  country  which  has  over  400  million  cattle,  is  the 


613 


assessment  of  the  extent  and  status  of  grasslands.  The  Space  Application  Center 
of  India  has  recently  undertaken  the  work  of  mapping  the  grasslands  of  the 
country  by  use  of  the  satellite  data.  The  work  is  yet  to  be  completed.  However, 
Lai,  Gulati  and  Bist  (1991)  did  a  limited  exercise  of  mapping  the  alpine  pastures 
in  the  Himalayas.  The  results  of  the  exercise  are  summarised  in  Table  3. 


Table  3 


Name  of  the  State 

Geographical  area 

Area  of  alpine 

pasture 

( in  million  ha) 

Arunachal  Pradesh 

8.36 

1.23 

Sikkim 

0.73 

0.16 

Uttar  Pradesh 

29.44 

0.85 

Himachal  Pradesh 

5.57 

1.73 

Jammu  and  Kashmir 

22  22 

13.19 

Total 

66  32 

17.16 

Lai  et  al  (1991)  assessed  the  interpretation  accuracy  by  ground-truth  verification 
of  sample  points.  The  number  of  sample  points  was  worked  out  by  using  the 
formula 


n=  p  x  q 

d2 

where  n=n umber  of  sample  points 

p=desired  proportion  of  estimated  accuracy 
q=(100-p) 

d=standard  error  acceptable  at  desired  confidence  level 
p,q,d  are  all  expressed  in  per  cent. 

The  sample  points  were  selected  by  random  sampling  out  of  a  population  formed 
by  12  x  12  cm  grids  on  the  map  of  alpine  pastures.  Though  it  was  beyond 
expectation,  the  interpretation  accuracy  was  found  to  be  100%. 


614 


UTILITY  OF  FOREST  VEGETATION  MAPS 


Apart  from  being  of  great  value  in  strategic  planning  at  the  national  level,  forest 
vegetation  maps  prepared  from  the  satellite  data  have  been  used  in  tactical 
planning  at  the  state  level.  From  the  third  assessment  "(FSI,1993)"  onward  the 
forest  vegetation  maps  have  been  prepared  district-wise.  A  district  is  the 
fundamental  administrative  unit  in  India,  and  forest  vegetation  maps'  being 
available  every  two  years  for  a  district  would  facilitate  operational  planning. 

Forest  vegetation  maps  obtained  from  satellite  data  gives  information  not  only 
on  change  in  land  cover,  but  also  on  change  in  land  use.  In  1989,  Forest  Survey 
of  India  "(Lai,  1989b)"  carried  out  an  exercise  to  assess  the  change  in  area  affected 
by  shifting  cultivation  in  north-eastern  hill  states  of  India.  The  exercise  indicated 
that  during  the  period  1974  to  1984,  area  affected  by  shifting  cultivation  decreased 
from  73410  ha  to  62854  ha.  The  State  of  Forest  Report,  1993  "(FSI,1994)"  makes 
a  useful  analysis  of  the  losses  and  gains  in  forest  cover  in  the  north-east  hill  states 
of  India  due  to  change  in  land  use.  The  report  points  out  that  there  was  a  loss  of 
70200  ha  of  forest  cover  on  account  of  more  forest  area  being  brought  under 
shifting  cultivation,  while  there  was  a  gain  of  31500  ha  in  forest  cover  on  account 
of  natural  regeneration  of  abondoned  shifting  cultivation  sites. 

CONCLUSIONS 

The  utility  of  satellite  data  in  land-cover  and  land-use  assessment  and  monitoring 
has  been  proved  beyond  doubt.  The  vegetation  maps  prepared  from  satellite  data 
are  fairly  accurate  as  the  emperical  testing  would  indicate,  but  a  fool-proof 
statistical  design  for  accuracy  assessment  has  yet  to  be  evolved  for  Indian 
conditions  which  are  ecologically  so  diverse.  For  interpretation  accuracy,  two 
points  have  to  be  specially  kept  in  mind  First,  in  regard  to  decidous  forest  it  has 
to  be  ensured  that  imagery  pertaining  to  the  period  trees  are  leafless  to  any 
degree  is  not  selected  for  interpretation  .  Secondly,  elaborate  experiments  have 
to  be  organised  to  test  whether,  as  Lai  has  pointed  out  "(Lai,  1989b)",  reflectance 
recorded  by  satellite  sensors,  is  more  a  function  of 'leaf  area  duration'  than  of 'leaf 
area  index'. 

REFERENCES 

Forest  Survey  of  India  (FSI).  1988.  State  of  Forest  Report,  1987.  7-20pp. 
Forest  Survey  of  India.  1989.  State  of  Forest  Report,1989.  7-1 1pp. 
Forest  Survey  of  India.  1993.  State  of  Forest  Report,  1991.  7-1 1  pp. 
Forest  Survey  of  India.  1994.  State  of  Forest  Report,  1993.  7-1 1pp. 
Lal,J.B.  1989a.  India's  Forests  :  Myth  and  Reality,  2nd  ed.  Dehradun  : 
Natraj  publisher  44,89-97  pp 


615 


LalJ.B.  1989b.  Forest  Resource  Management  from  Space.  Van  Vigyan. 

Vol.  27,No.4.  1989.  168-178pp. 
LalJ.B.,  and  R.C  Prajapati.  1989.  Effect  of  Scale  on  Vegetation  Mapping 

be  Remote  Sensing.  The  Tndian  Forester.  Vol.  115,  No.  6,  June,  1989. 

372-377pp. 

LalJ.B,  A.K.  Gulati  and  M.S.  Bist.1991.  Satellite  mapping  of  alpine 
pastures  in  the  Himalayas.  Tnt.  J.Remote  Sensing.  1991,  Vol.  12,  No.  3. 
435-443pp. 

BTOGRAPHTCAL  SKETCH 

J.B.  Lai,  a  former  member  of  Tndian  Forest  Service,  is  an  Associate  of  Indian 
Forest  College,  and  a  Ph.D  from  the  Victoria  University,  Manchester,U.K.  He  has 
held  top  forestry  positions  in  the  Govt,  of  India.  He  was  Director,  Forest  Survey 
of  India,  Director  General,  Indian  Council  of  Forestry  Research  and  Education, 
and  Director,  Indian  Institute  of  Forest  Management.  He  has  served  as  a  member 
of  National  Enviromental  Council  of  India,  and  as  a  consultant  to  FAO  and  STDA. 
He  has  written  three  books,  and  several  technical  papers. 


616 


Study  About  the  Use  of  SPOT 
Data  for  the  Estimation  of  Pastoral 
Areas  in  Morocco 


Philippe  Brion1 


Abstract. -In  order  to  provide  information  on  land  use  for  a  whole 
country,  it  is  often  impossible  to  have  a  complete  cover  of  the  land  with 
remote  sensing  data,  and  it  is  necessary  to  use  a  sampling  method.  This 
paper  presents  an  approach  of  this  problem,  relatively  to  the  estimation 
of  pastoral  areas  in  Morocco  :  selection  of  SPOT  images,  and  then 
selection  of  "segments"  inside  the  images.  The  different  options  of  the 
sampling  plan  are  discussed. 


INTRODUCTION 

In  order  to  provide  information  on  land  use,  remote  sensing  brought  many 
hopes.  In  some  cases,  it  has  been  used  jointly  with  ground  surveys,  and  the 
exhaustive  cover  of  the  satellite  image  has  given  an  auxiliary  information, 
combined  with  the  results  of  the  ground  survey  in  order  to  produce  an  improved 
estimate  (by  the  method  of  the  regression  estimate,  see  for  example  Pastorelli, 
1992).  But,  often,  the  number  of  images  required  for  having  an  exhaustive  cover  is 
too  high,  and  financial  reasons  lead  to  the  use  of  a  sample  of  images  :  this  is,  for 
example,  the  case  of  the  MARS  program  (of  the  European  commission,  for  the 
estimation  of  crop  areas  as  cereals)  ;  however,  there  are  no  established  rules  for 
the  sampling  plan  to  be  used  for  that  kind  of  problems. 

The  object  of  this  paper  is  to  propose  an  a  priori  approach  to  this  problem, 
concerning  estimation  of  pastoral  areas  in  Morocco.  This  work  was  done  jointly 
with  CNES  (Centre  National  d'Etudes  Spatiales,  Toulouse)  and  CRTS  (Centre 
Royal  de  Teledetection  Spatiale,  Rabat). 


Statistician,  IS'SEE.  Pans.  France 


617 


PRINCIPLES  OF  THE  PROPOSED  APPROACH 


Constraints  and  method 

The  first  constraint  concerns  the  purchase  of  satellite  data  (here,  SPOT  images). 
The  impossibility  to  have  a  complete  cover  of  Morocco  leads  us  to  have  a  sample 
of  images.  Then,  since  the  classification  of  images  does  not  presently  give  quite 
perfect  results  (there  still  remains  confusions  for  some  types  of  land  use),  we  are 
led  to  having  more  complete  investigations  for  portions  of  the  images  (by 
photointerpretation  for  example). 

So,  we  propose  a  two-stage  sampling  plan  for  the  estimation  of  pastoral  areas: 

-  first,  selection  of  SPOT  images  ; 

-  second,  selection  of  "segments"  inside  the  selected  images. 

These  segments  are  "surveyed"  by  photointerpretation,  or  by  a  ground  survey. 
They  have,  for  practical  reasons,  a  square  shape. 

Which  items  of  the  sampling  plan  are  to  be  studied  ? 

-  The  size  of  the  segments  has  to  be  chosen.  In  France,  for  example,  a  size  of  50 
hectares  is  used  for  ground  surveys  because  of  practical  reasons  (work  of  the 
enumerator).  What  should  be  proposed  in  the  context  of  pastoral  areas  in 
Morocco? 

-  Another  question  concerns  the  number  of  segments  to  be  selected  inside  a  SPOT 
image. 

-  Third,  the  sample  of  segments  may  be  selected  using  a  systematic  sampling 
method.  Is  this  method  efficient,  and  if  so,  what  is  the  gain  of  precision  to  be 
expected  ? 

-  The  selection  of  SPOT  images  is  another  problem  :  how  many  should  be 
selected,  and  is  it  possible  to  use  a  stratification  of  the  country  for  the  sampling 
plan? 

Then,  once  the  former  points  have  been  treated,  what  global  precision  is  to  be 
expected  ? 

An  experimental  approach 

This  approach  consists  in  using  some  existing  materials  in  order  to  produce,  a 
priori,  some  quantified  results  to  the  former  questions. 

First,  three  SPOT  images,  belonging  to  regions  of  Morocco  with  different 
characteristics  (in  the  upper  Atlas,  near  the  Sahara  and  on  arid  plateaux)  have  been 
classified,  especially  isolating  the  item  "pastoral  area"  ;  these  images  have  been 
used  for  the  study  of  the  second  stage  of  the  sampling  plan. 

The  study  of  the  first  stage  was  realized  using  data  proceeding  from  a  SPOT 
"Quick-Look-Numerical"  (QLN)  :  these  data  are  obtained  by  sampling  one  pixel 
for  six,  on  row  and  column,  and  give  a  global  view  of  the  country,  much  bigger 


618 


than  a  single  image.  These  data  were  also  classified,  particularly  using  the  item 
"pastoral  area". 

The  use  of  these  materials  permits  to  have  quantified  data  to  study  the  sampling 
system  before  the  actual  survey  :  this  approach  is  rather  a  new  one  for  the  problem 
studied,  because  until  now,  we  have  had  at  our  disposal  experiences  where  the 
precision  is  evaluated  a  posteriori,  once  the  survey  has  been  conducted. 

At  last,  it  is  also  important  to  mention  that  the  subject  discussed  in  this  paper  is 
a  statistical  one,  and  not  a  cartographic  one. 

RESULTS 

First,  we  will  discuss  the  results  concerning  the  second  stage  (selection  of 
segments  inside  the  images),  and  integrate  them  to  those  obtained  for  the  first 
stage  (selection  of  SPOT  images). 

Sample  of  segments  inside  the  images 

The  three  images  classified  were  cut  up  in  segments,  using  six  different  grids  to 
define  the  segments  :  200*200m,  400*400m,  600*600m,  800*800m, 
1000*  1000m,  1200*  1200m. 

Size  of  the  segments 

The  variance  of  the  estimate  of  the  mean  of  a  variable  Y,  using  a  simple  random 
sample  of  m  segments  is  approximately  equal  to  : 

V (estimate  of  the  mean)  =  V(Y)  /  m  (1) 

if  V(Y)  is  the  variance  of  the  variable  in  the  population,  and  if  we  do  not  take 
into  account  the  sampling  fraction  /  that  is  very  small,  so  the  quantity  (1-j)  is 
approximately  equal  to  1 . 

Table  1  displays  the  value  of  V(Y)  for  the  variable  "proportion  of  pastoral  areas 


inside  the  segments" 

,  regarding  the  three 

images  and  the 

different  sizes 

segments. 

Table  1.-  V(proportion 

of  pastoral  areas  inside  the  segments) 

Segment  size 

Midelt 

Oujda 

Ktaoua 

200  *  200 

0.132 

0.078 

0.165 

400  *  400 

0.111 

0.067 

0.144 

600  *  600 

0.102 

0.059 

0.128 

800  *  800 

0.095 

0.057 

0.1 17 

1000  *  1000 

0.091 

0.054 

0.108 

1200  *  1200 

0.087 

0.050 

0.100 

When  the  size  of  the  segment  increases,  the  variance  V(Y)  decreases,  because  the 
bigger  segments  are  more  homogeneous.  But,  if  we  introduce  cost  elements 


619 


(corresponding  to  the  time  needed  to  "survey"  the  segments,  e.g.  proportional  to 
the  area  of  the  segments,  or  at  least  to  the  square  root  of  this  area),  we  see  that  the 
decrease  of  variance  is  not  sufficient  enough  to  compensate  cost  increase.  So, 
according  to  the  results  obtained  for  these  three  images  located  in  different  regions 
of  Morocco,  it  seems  better  to  use  small  segments  (200*200m). 

Efficiency  of  the  systematic  sampling  method 

This  method  consists  in  selecting  segments  located  on  a  grid,  rather  than  taking 
them  completely  randomly.  The  size  of  the  grid  may  vary,  and  is  related  to  the 
sampling  fraction.  Having  cut  up  three  images  in  segments  makes  it  possible  to 
calculate  exactly  the  variance  of  the  estimate  with  this  method  (by  calculating  the 
results  of  all  possible  systematic  samples,  which  are  in  limited  number),  and  to 
compare  it  to  the  one  obtained  by  Eq.  1,  for  the  three  sites,  different  sizes  of 
segments  and  for  different  sampling  rates. 

The  results  are  that,  generally,  the  systematic  method  is  more  efficient  that 
simple  random  sampling,  and,  on  an  average,  the  gain  of  variance  of  the  estimate 
is  between  30%  and  40%.  These  results  are  similar  to  those  obtained  by  Dunn  and 
Harrison  (1993). 

Precision  of  the  estimation  inside  the  images 

Table  2  gives  the  coefficient  of  variation  for  the  estimate  of  the  proportion  of 
pastoral  areas  inside  each  image  for  a  systematic  sampling  of  segments  of 
200*200m,  the  sampling  fraction  being  1/100. 


Table  2.-Precision  of  the  estimates  inside  the  images 


Image 

coefficient  of  variation 

Midelt 

2.3  % 

Oujda 

1.1  % 

Ktaoua 

9.9  % 

The  coefficient  of  variation  is  defined  as  the  ratio  from  the  square  root  of  the 
variance  of  the  estimate  of  a  characteristic  to  the  value  of  this  characteristic. 


Selection  of  SPOT  images 

Using  the  QLN  makes  it  possible  to  reconstruct  "approximate"  SPOT  images  for 
the  whole  country.  The  values  obtained  from  these  images  constitute  the  basis  to 
evaluate  the  global  precision  of  the  sampling  plan. 

Variance  of  the  estimate  for  a  two-stage  sampling  plan 

If  the  selection  for  the  first  stage  is  with  equal  probabilities,  the  variance  of  the 

estimate  Y  of  the  total  is  : 


620 


m  v     m  J  m 

7;  ( Y)    pastoral  area  of  the  image  a 

Za  is  the  variance  of  the  estimate  of  the  total 

Ta  (  Y)  according  to  the  method  of  selection 

of  segments  inside  the  image 


Results  from  the  first  part  (selection  of  segments  inside  the  images)  are  used  for 
the  estimation  of  the      ,  and  the  data  of  QLN  give  information  about  Sf . 

Results 

Average  results  (calculated  with  the  three  images)  for  the  second  stage  were 
used,  and  Table  3  gives  the  precision  of  the  global  estimate  of  pastoral  areas  for 
Morocco,  with  different  numbers  of  images  selected  at  the  first  stage,  and  400 
segments  of  200*200m  selected  in  each  image. 

Table  3. -Precision  of  the  estimate  of  pastoral  area  for  Morocco  


Number  of  surveyed  images       Coefficient  of  variation  of     Relative  share  of  the  second 

the  estimate  (%)           stage  in  the  total  variance  of 
 the  estimate  (%)  


20 

12.0 

2.1 

30 

9.3 

2.4 

40 

7.6 

2.7 

50 

6.3 

3.1 

70 

4.5 

4.3 

100 

2.2 

12.3 

The  share  of  the  variance  which  depends  from  the  second  stage  is  small,  that 
means  that  the  variance  between  images  brings  the  most  important  part  of  this 
variance.  So,  it  is  possible  to  have,  for  the  second  stage,  a  smaller  sample.  If  we 
take  30  images  and  30  segments  inside  each  image,  we  obtain  a  coefficient  of 
variation  of  10.5%. 

The  second  result  of  Table  3  is  that,  even  when  the  sampling  rate  for  the  first 
stage  is  high,  for  example  100  images  selected  within  the  115  "possible"  which 
cover  the  whole  country,  there  still  remains  a  consistent  value  of  the  variance  of 
the  estimate. 


621 


Use  of  a  stratification 

A  "rough"  stratification  was  used,  classifying  all  115  images  in  9  classes  : 
selecting  30  images  using  this  stratification  (and  30  segments  inside  each  image), 
the  coefficient  of  variation  of  the  estimate  for  pastoral  areas  is  8.9%  (instead  of 
10.5%  without  stratification). 

CONCLUSION 

The  most  important  result  of  this  study  is  the  great  part  of  the  variance  of  the 
estimate  due  to  the  first  stage  of  the  sampling  plan,  that  is  the  fact  of  selecting 
images  among  the  115  images  covering  the  country. 

This  result  is  well  known  for  sampling  questions,  and  one  way  to  take  it  into 
account  and  to  improve  the  efficiency  of  the  method  might  be  to  have  smaller 
primary  units  :  using  portions  of  SPOT  images  for  the  first  stage  could  be  a 
response  to  that  problem. 

Then,  the  results  obtained  in  this  paper  should  not  be  considered  as  exact  values 
of  the  precision  to  be  obtained  with  such  a  method,  but  rather  as  approximations 
based  on  an  experimental  approach. 

Finally,  we  considered  this  problem  as  an  "instantaneous"  estimation,  trying  to 
evaluate  the  pastoral  areas  of  the  country  for  one  year.  To  evaluate  the  evolution 
between  two  years,  or  more,  should  lead  to  other  propositions  and  results  for  the 
sampling  plan. 

REFERENCES 

Cochran,  W.G.,  1977,  Sampling  techniques,  3rd  ed,  Wiley,  New  York 

Dunn,  R.,  Harrison,  A.R.,  1993,  Two-dimensional  systematic  sampling  of  land 

use,  Applied  Statistics  42  n°4 

Pastorelli,  R,  1992, Superficies  agricoles  a  partir  d'images  satellite,  Courrier  des 

statistiques  n°61-62,  Paris 
Rapport  CNES,  1995,  Projet  Geostat  Maroc,  etude  statistique,  CNES,  Toulouse 

BIOGRAPHICAL  SKETCH 

Philippe  Brion  graduated  from  Ecole  Polytechnique  (Paris)  and  ENSAE  (Ecole 
Nationale  de  la  Statistique  et  de  1' Administration  Economique,  Paris).  He  was  in 
charge  of  the  sampling  methodology  in  SCEES  (statistical  department  of  the 
French  Ministry  of  Agriculture)  from  1980  to  1989,  and  is  now  working  in  INSEE 
(Institut  National  de  la  Statistique  et  des  Etudes  Economiques)  on  methodological 
problems  for  statistical  cooperation  (developing  and  transition  countries). 


622 


The  Influence  of  Vegetation  Cover 
Density  and  Topographic  Parameters 

on  the  Thermal  Emission  of  the 
Beech  Forests  of  Simbruini  Mountains 

(Central  Italy) 

Carlo  Ricotta,  Giancarlo  Avena  and  Fernando  Ferri1 


Abstract.  -  Numerous  studies  have  noted  a  strong  negative  correlation 
between  radiometric  surface  temperature  and  spectral  ratio-based 
vegetation  indices.  Since  vegetation,  through  the  process  of 
transpiration,  considerably  enhances  evaporative  fluxes  to  the 
atmosphere,  greater  vegetation  biomass  should  be  associated  with 
increased  reduction  in  surface  temperature.  In  irregular  terrains  with 
variable  sun-sensor-surface  geometry  one  of  the  major  problems 
concerning  the  evaluation  of  surface  temperatures  of  vegetation  from 
remotely  sensed  data  is  that  the  thermal  response  of  the  earth  surface  is 
strongly  influenced  by  topographic  factors,  such  as  altitude  and 
insolation  angle.  In  this  paper,  which  uses  remotely  sensed  inputs  of 
surface  temperature  and  vegetation  fractional  cover,  the  relations 
between  the  surface  temperatures  of  the  beech  forests  of  Simbruini 
Mountains  (Central  Italy)  and  the  main  vegetation  cover  density  and 
topographic  parameters  are  discussed.  According  to  the  results  obtained, 
79.0%  of  the  thermal  response  variance  of  the  beech  forests  is  due  to 
variations  of  the  topographic  parameters. 


INTRODUCTION 

Thermal  infrared  measures  of  the  Earth's  surface  from  terrestrial  and  airborne 
sensors  have  been  used  since  the  1 960s  to  investigate  energy  budget  conditions  of 
vegetation.  With  the  launch  of  Landsat  4  and  5  plant  ecologists  now  have  access 
to  a  wide-area  synoptic  and  multitemporal  thermal  infrared  band  TM6  with  a 

Remote  Sensing  Laboratory,  Department  of  Plant  Biology,  University  of  Rome  "La  Sapienza",  Rome,  Italy. 


623 


relatively  high  resolution  (Lathrop  e  Lillesand  1987).  A  number  of  authors  (Asrar 
et  al.  1988,  Smith  and  Choudhury  1990,  Smith  and  Choudhury  1991,  Hope  and 
McDowell  1992)  had  focused  on  the  widely  observed  negative  correlation 
between  radiometric  surface  temperature  and  remotely  sensed  measures  of 
actively  transpiring  vegetation  such  as  vegetation  indices. 

One  hypothesis  to  explain  this  observation  is  that  variability  in  transpiring  leaf 
area  produces  spatial  variability  in  surface  temperature  by  modulating  evaporative 
cooling  at  the  land-surface-atmosphere  interface  (Friedl  and  Davis  1994).  In 
mountainous  regions,  however,  this  relationship  may  be  complicated  by  the  effect 
of  topography  since  a  variation  in  altitude  and  insolation  angle  should  also  result 
in  variated  surface  heating.  The  combined  effect  of  topography  and 
evapotranspiration  in  the  energy  exchange  process  at  the  surface  produces  a 
complex  problem  in  the  interpretation  of  observed  thermal  emissions  (Conese  et 
al.  1989).  The  present  work  aims  at  estimating  by  Landsat  5  TM  data  the 
variability  of  the  beech  forest  surface  temperatures  located  on  the  Simbruini 
Mountains  (Central  Italy)  depending  on  the  major  remotely  sensed  vegetation 
cover  density  and  topographic  parameters,  such  as  the  Normalized  Difference 
Vegetation  Index  (NDVI),  altitude  and  insolation  angle. 

STUDY  AREA 

The  beech  forests  of  the  Municipalities  of  Trevi  nel  Lazio  and  Filettino 
(Central  Italy)  were  selected  for  analysis.  There  are  dense  5600  ha  forests  located 
in  the  watershed  of  the  Aniene  River  with  elevational  extent  from  about  900  to 
1700  m.  Due  to  environmental  importance,  the  area  was  designated  in  1983  as 
part  of  the  Natural  Park  of  Simbruini  Mountains. 

The  study  area  belongs  to  the  Mountainous/Sub-alpine  bio-climatic  belt  (Blasi 
1994).  The  climate  is  strongly  influenced  by  the  position  of  the  Simbruini 
Massive,  which  forms  a  transversal  barrier  to  the  western  sea- winds. 
Meteorological  data  came  from  the  station  of  the  Ministero  dei  Lavori  Pubblici 
located  in  Filettino  (1062  metres  above  sea  level.).  The  average  annual 
temperature  is  10.7  °C.  The  average  annual  rainfall  is  1413.9  mm  with  a  principal 
maximum  in  winter  and  a  secondary  maximum  during  fall.  The  average  rainfall  in 
the  quarter  June-August  is  155.7  mm  with  a  consequent  lack  of  summer  aridity.  A 
short  period  of  summer  sub  aridity  is  still  possible  during  particularly  dry  years. 
The  range  of  relief  is  approximately  1 ,600  metres. 

The  maximum  elevation  above  sea  level  is  2,156  metres  (MSL).  The 
topography  of  the  area  is  complex  with  a  variety  of  site  exposures.  The  beech 
forests  represent  a  nearly  continuous  formation  with  high  cover  density  values  due 
to  the  absence  of  intense  grazing  and  cutting  activities.  The  dominant  tree  species 
is  Fagus  sylvatica.  More  rare  arboreal  species  are  Acer  pseudoplatanus,  Taxus 
baccata,  Tilia  platiphyllos  and  Ulmus  glabra. 


624 


MATERIALS  AND  METHODS 


The  study  area  was  extracted  from  a  Landsat  TM  190/31  II  Q  scene  from  29 
July  1993.  Collateral  information  consisted  of  the  vegetation  map  of  the  study 
area  at  1:25,000  scale  obtained  through  photointerpretation  of  infrared  aerial 
photographs.  Minimal  and  maximal  air  temperatures  were  respectively  12.8  °C 
and  26.8  °C  at  Filettino  on  29  July  1993.  In  the  month  prior  to  satellite  overpass 
there  was  79.4  mm  of  rain.  17  July  was  the  last  rainy  day  prior  to  satellite 
overpass  with  70.0  mm  of  rain.  On  29  July  1993  the  atmosphere  was  clear  and 
any  atmospheric  attenuation  was  assessed  to  be  systematic  across  the  study  area. 

The  TM  data  were  rectified  to  a  Universal  Transverse  Mercator  grid  using  a 
nearest  neighbour  resampling  method.  A  digital  terrain  model  (DTM)  was  created 
by  acquiring  on  a  digitizer  the  contour  lines  every  25  m  of  the  national 
cartography  mapping  sheets  at  1:25,000  scale  of  the  study  area.  Relative  sun 
incidence  was  calculated  from  the  DTM.  Relative  sun  incidence  displays  direct 
illumination  which  is  equal  to  the  cosine  of  the  angle  between  the  surface  normal 
and  the  incident  beam  for  each  pixel.  The  solar  azimuth  and  altitude  angles  at  the 
time  of  satellite  overpass  (10:09  local  time)  necessary  to  calculate  this 
illumination  model  were  read  from  the  header  file  of  the  computer  compatible 
tape  (CCT). 

In  order  to  test  the  hypothesis  that  thermal  emissions  are  related  to  the 
topographic  parameters,  such  as  altitude  and  insolation  angle  and  the  amount  of 
vegetative  cover,  a  systematic  sampling  (Congalton  1988)  of  216  pixels  was 
performed  on  the  vegetated  areas  of  the  image.  To  avoid  autocorrelation  effects 
particularly  on  band  TM  6  the  starting  point  was  located  at  random  and  each 
successive  unit  was  taken  at  an  interval  of  10  lines  and  10  columns  thereafter. 

For  each  of  those  pixels  the  values  of  altitude  (MSL),  relative  sun  incidence 
(COS  Z),  NDVI  (TM4-TM3/TM4+TM3)  and  surface  temperatures  were  extracted 
from  the  data  set. 

THERMAL  DATA  CALIBRATION 

The  satellite-derived  normalized  difference  vegetation  index  (NDVI)  is  used  as 
the  index  for  the  vegetation  density.  The  physical  quantity  that  is  measured 
remotely  is  the  spectral  radiance  reaching  the  sensor  [mW/cm^  sr  urn]  quantized 
in  digital  numbers  (DN)  over  a  8-bit  brightness  range.  Digital  counts  of  the  bands 
TM3,  TM4  and  TM6  were  converted  to  spectral  radiances  according  to  the 
method  proposed  by  Markham  and  Barker  (1986). 

The  improved  darkest  object  subtraction  technique  described  by  Chavez  (1988) 
was  used  to  correct  digital  numbers  from  the  red  and  near-infrared  bands  for 
differences  in  atmospheric  scattering  in  these  two  bands.  This  simple  correction 
was  applied  to  minimize  compression  in  the  NDVI  values  which  results  from 


625 


unequal  atmospheric  effects  on  the  red  and  near-infrared  radiances  (Hope  and 
McDowell  1992).  Spectral  radiance  L  in  the  thermal-infrared  band  and  surface 
temperature  Ts  are  related  in  the  form  (Singh  1988): 

ln(L)  =  a  +  b/Ts  [1] 

Note  that  the  accuracy  of  Eq.  1  decreases  with  the  increase  of  the  range  of 
temperature  chosen.  For  this  reason  the  parameters  a  and  b  were  determined  for 
temperature  range  of  40  °K  only  and  are  reported  by  Singh  (1988). 

This  technique  does  not  account  for  atmospheric  attenuation  of  the  thermal 
radiance.  However,  although  the  8-14  urn  atmospheric  window  is  far  from  being 
transparent,  Landsat  TM  thermal  IR  data  collected  under  clear  atmospheric 
conditions  can  be  correctly  calibrated  in  surface  temperatures  with  acceptable 
accuracy  even  when  corrections  for  atmospheric  effect  are  not  applied.  This  is 
possible  because  under  clear  atmospheric  conditions  the  amount  of  atmospheric 
path  radiance  is  nearly  equal  to  the  amount  of  attenuated  target  radiance 
(Bartolucci  et  al.  1988).  Moreover,  since  the  objective  of  this  study  was  to 
investigate  the  correlation  coefficients  of  the  relationships  between  Ts  and  the 
examined  independent  variables,  a  systematic  offset  in  the  values  would  not  affect 
the  interpretation  of  the  results.  The  relationship  between  the  surface  temperature 
of  the  beech  forests  (Ts)  and  the  three  independent  variables  MSL,  COS  Z  and 
NDVI  were  investigated.  Linear  correlation  coefficients  were  calculated  to 
quantify  the  magnitude  of  the  hypotesized  relationships,  and  scattergrams  were 
constructed  to  indicate  the  nature  of  the  relationships. 

RESULTS  AND  DISCUSSION 

Scattergrams  depicting  the  relationship  between  surface  temperatures  of  the 
beech  forests  and  the  independent  variables  are  presented  in  figure  1  and  2,  and 
the  associated  correlation  coefficients  are  given  in  table  1 . 


Table  1.  -  Correlation  coefficients  (R)  and  coefficients  of  determination  (R2)  for  surface 
temperatures  of  vegetation  versus  values  from  MSL,  COS  Z  and  NDVI.  *)  Not  significant  at 
the  0.001  level. 


MSL 

cosz 

NDVI 

R 

-0.606 

0.640 

R2 

0.367 

0.410 

* 

626 


These  linear  correlation  coefficients  show  that  at  time  of  satellite  overpass 
COS  Z  has  the  strongest  association  with  surface  beech  forests  temperatures.  This 
fact  can  be  explained  with  the  obvious  warming  of  vegetation  surfaces  caused 
after  sunrise  on  illuminated  slopes.  It  is  very  likely  that  at  the  time  of  satellite 
overpass  (10:09  local  time)  the  temperature  difference  between  illuminated  and 
shaded  surfaces  is  near  its  daily  maximum  (Conese  et  al.  1989).  On  29  July  1993, 
the  relationship  between  MSL  and  the  surface  temperature  of  the  beech  forests 
was  weaker  with  a  negative  correlation  coefficient.  Moreover,  including  MSL 
along  with  COS  Z  in  the  calculation  of  a  bivariate  correlation  coefficient  did 
explain  79%  of  the  variance  of  the  surface  temperature  of  the  beech  forests  (table 
2).  On  the  contrary,  according  to  Smith  and  Choudhury  (1991),  the  simple  linear 
regression  of  the  surface  temperature  of  the  beech  forests  with  the  corresponding 
remotely  sensed  vegetation  index  amount  had  no  significant  correlation 
coefficient.  In  fact,  no  strong  inverse  relationships  between  surface  temperatures 
and  vegetation  cover  density  indexes  are  evident  for  homogeneous  forests  with 
complete  ground  cover.  This  relationships  exist  only  for  crop  and  pasture  lands 
where  fractional  vegetation  cover  varies  from  bare  soil  trough  to  crops  with 
complete  ground  cover  giving  a  wide  range  of  surface  temperature  and  NDVI 
values. 


Table  2.  -  Bivariate  and  multivariate  coefficients  of  determination  (R2)  for  surface 
temperatures  of  vegetation  versus  values  from  MSL,  COS  Z  and  NDVI.  All  coefficients  are 
significant  at  the  0.001  level. 


MSL 

MSL 

MSL 

NDVI 

COSZ 

COSZ 

NDVI 

COSZ 

NDVI 

R2 

0.790 

0.374 

0.413 

0.792 

A  week  relationship  between  NDVI  and  surface  temperature  of  the  beech 
forests  is  still  evident  in  combination  with  the  topographic  variables.  The  multiple 
correlation  coefficient  of  the  surface  temperature  of  the  beech  forests  with  the 
three  independent  variables  MSL.  COS  Z  and  NDVI  increase  their  explained 
variance  by  less  than  1%  with  respect  to  the  bivariate  correlation  coefficient  of  the 
topographic  variables  alone  with  surface  temperature. 

CONCLUSIONS 

This  study  has  defined  some  of  the  problems  encountered  when  attempting  to 
use  remote  sensed  data  to  estimate  surface  temperatures  of  vegetation  in  complex 


627 


vegetated  mountainous  regions.  In  irregular  terrains  the  influence  of  topography 
can  strongly  affect  the  thermal  levels  of  vegetation.  The  comprehension  of  the 
sources  of  this  influence  is  thus  determinant  for  a  correct  interpretation  of 
remotely  sensed  thermal  data. 

Results  of  the  correlation  analysis  on  29  July  1993  in  mid-summer  sub-aridity 
conditions  reveal  that  at  time  of  satellite  overpass  COS  Z  alone  accounted  for 
41.0%  of  the  variance  in  surface  temperature  of  the  beech  forests.  This  results 
indicate  that  insolation  angle  was,  under  the  conditions  of  this  experiment,  a  more 
important  factor  controlling  surfaces  temperatures  of  the  beech  forests  than 
altitude  or  vegetation  cover  density.  Moreover,  the  combined  effect  of  differential 
heating  and  altitude  explains  79.0%  of  the  variance  of  surface  temperature  of  the 
beech  forests  of  Simbruini  Mountains.  On  the  contrary,  the  hypothesized  negative 
relationship  between  NDVI  and  surface  temperature  is  not  significant  due  to  the 
limited  range  of  surface  temperatures  and  fractional  vegetation  cover  density 
values  of  the  beech  forests  of  the  study  area. 

However,  remotely  sensed  data  provide  only  an  instantaneous  assessment  of 
thermal  conditions  and  the  use  of  remote  sensed  thermal  data  is  just  a  first  step  to 
describe  surface  temperatures  of  vegetation.  Future  works  towards  accurately 
define  thermal  trends  of  vegetation,  should  necessarily  include  field 
measurements  to  convert  instantaneous  satellite-derived  thermal  data  to  daily 
predictions. 


320 


310 


=>  300 

1—4 
U 


280  _|  

800  1100  1400  1700  2000 


Figure  1.  -  Scattergram  of  surface  temperatures  versus  MSL 


628 


320 


310 


300 


280  J  

' 0  0.25  0.50  0.75  1 

cos  z: 


Figure  2.  -  Scattergram  of  surface  temperatures  versus  COS  Z 


REFERENCES 

Asrar,  G.,  Harris,  T.R.,  Lapitan,  R.L.  and  Cooper,  D.I.  1988.  Radiative  surface 
temperature  of  the  burned  and  unburned  areas  in  a  tallgrass  prairie,  Remote 
Sensing  of  Environment,  24/447-457. 

Bartolucci,  L.A.,  Chang,  M.,  Anuta,  P.E.  and  Graves,  M.R.  1988.  Atmospheric 
effects  on  Landsat  TM  thermal  IR  data,  IEEE  Transactions  on  Geoscience  and 
Remote  Sensing,  26/171-176. 

Blasi,  C.  1994.  Fitoclimatologia  del  Lazio.  Regione  Lazio,  Ufficio  Parchi  e 
Riserve  Naturali,  Roma. 

Chavez,  P.S.  1988.  An  improved  dark-object  subtraction  technique  for 
atmospheric  scattering  correction  of  multispectral  data,  Remote  Sensing  of 
Environment,  24/459-479. 

Conese,  C  Maselli,  F.  and  Maracchi,  G.  1989.  Influence  of  orography  on  the 
thermal  response  of  coniferous  forests  measured  by  TM  data.  User  contribution 
to  satellite  remote  sensing  programmes,  Proceedings  of  the  IX  EARSEL 
Symposium  held  in  Espoo,  Finland,  on  27  June  1989,  pp.  267-273. 

Congalton,  R.G.  1988.  A  comparison  of  sampling  schemes  used  in  generating 
error  matrices  for  assessing  the  accuracy  of  maps  generated  from  remotely 
sensed  data.  Photogrammetric  Engineering  and  Remote  Sensing,  54/593-600. 


629 


Friedl,  M.A.  and  Davis,  F.W.  1994.  Sources  of  variation  in  radiometric  surface 
temperature  over  a  tallgrass  prairie,  Remote  sensing  of  Environment,  48/1-17. 

Hope,  A.S.  and  McDowell,  T.P.  1992.  The  relationship  between  surface 
temperature  and  a  spectral  vegetation  index  of  a  tallgrass  prairie:  effects  of 
burning  and  other  landscape  controls,  International  Journal  of  Remote  Sensing, 
13/2849-2863. 

Lathrop,  R.G.  and  Lillesand,  T.M.  1987.  Calibration  of  Thematic  Mapper  thermal 

data  for  water  surface  temperature  mapping:  case  study  on  the  Great  Lakes, 

Remote  Sensing  of  Environment,  22/297-307. 
Markaham,  B.L.  and  Barker,  J.L.  1986.  Landsat  MSS  and  TM  post-calibration 

dynamic  ranges,  exoatmospheric  reflectances  and  at-satellite  temperatures, 

Landsat  Technical  Notes,  1/3-8. 
Singh,  S.M.  1988.  Brightness  temperature  algorithms  for  Landsat  Thematic 

Mapper  data,  Remote  Sensing  of  Environment,  24/509-512. 
Smith,  R.C.G.  and  Choudhury,  B.J.  1990.  On  the  correlation  of  indices  of 

vegetation  and  surface  temperature  over  south-eastern  Australia,  International 

Journal  of  Remote  Sensing,  11/2113-2120. 
Smith,  R.C.G.  and  Choudhury,  B.J.  1991.  Analysis  of  normalized  difference  and 

surface  temperature  observations  over  southeastern  Australia,  International 

Journal  of  Remote  Sensing,  12/2021-2044. 

BIOGRAPHICAL  SKETCH 

Carlo  Ricotta  is  a  Ph.D.  candidate  in  Botanical  Sciences  at  the  Department  of 
Plant  Biology  of  the  University  of  Rome  "La  Sapienza",  Italy,  where  he  graduated 
in  Geology  in  1992.  His  research  interests  include  the  extraction  of  spatial 
information  from  remotely  sensed  imagery  to  study  vegetation  dynamic  and 
landscape  change. 

Giancarlo  Avena  is  Full  Professor  of  Phytogeography  and  Director  of  the 
Department  of  Plant  Biology  at  the  University  of  Rome  "La  Sapienza".  He 
graduated  in  Geology  from  the  University  of  Rome  "La  Sapienza"  in  1962  and 
specialized  in  Physical  Geography  and  Geobotany.  His  current  research  interests 
include  the  application  of  remote  sensing  and  associate  techniques  for  the  study  of 
vegetation. 

Fernando  Ferri  is  technical  co-ordinator  of  the  Remote  Sensing  Laboratory  of 
the  Department  of  Plant  Biology  at  the  University  of  Rome  "La  Sapienza",  where 
he  graduated  in  Geology  in  1981.  Mr.  Ferri  provides  technical  consultation  on 
GIS  and  remote  sensing  projects  that  affect  vegetation  mapping. 


630 


Application  of  Non-Parametric  Kernel 
Regression  and  Nearest-Neighbor 
Regression  for  Generalizing  Sample 

Tree  Information 

Annika  Kangas1  and  Kari  T.  Korhonen2 


Abstract  -  Usually  in  a  forest  survey,  large  part  of  the  measured  trees 
are  tally  trees  from  which  only  elementary  characteristics  are  measured. 
Part  of  the  trees  are  sample  trees,  which  are  measured  more  thoroughly. 
To  be  able  to  utilize  tally  trees  efficiently  in  the  calculations  the 
information  available  for  sample  trees  has  to  be  generalized  for  tally 
trees.  The  generalized  information  should  be  unbiased  for  areas  of 
arbitrary  size  (or  arbitrary  groups  of  data).  Also,  the  variation  between 
sample  plots  and  within  sample  plots  should  be  realistic,  if  this  data  are 
used  as  a  basis  for  simulations  of  forest  development.  These  requirements 
may  often  be  contradictory.  In  this  paper  applications  of  non-parametric 
kernel  regression  and  nearest-neighbor  regression  for  generalizing  sample 
tree  information  are  discussed.  These  methods  may  provide  a  satisfactory 
compromise  between  these  requirements. 


INTRODUCTION 

Two-  or  multi-phase  sampling  is  applied  in  most  forest  inventory  systems.  The 
first  phase  sample  (tally  trees)  consists  of  a  large  number  of  trees  for  which 
diameter  and  other  easily  measurable  characteristics  are  measured.  The  second 
phase  sample  (sample  trees)  consists  of  trees  measured  more  thoroughly.  Height, 
age  and  additional  diameters  are  the  most  typical  characteristics  measured  for  the 
sample  trees.  A  third  phase  sample  may  be  collected  to  derive  volumes  or 
biomass  from  sample  tree  characteristics.  (Cunia  1986). 

If  two  phase  sampling  is  applied,  the  sample  tree  information  has  to  be 
generalized  for  tally  trees.  This  means  that  for  every  tally  tree  an  expected  value 


'  Research  Scientist,  Finnish  Forest  Research  Institute,  Kannus  Research  Station,  P.O.  BOX  44,  FIN-60101  Kannus, 
Finland 

2  Research  Scientist,  Finnish  Forest  Research  Institute,  Joensuu  Research  Station,  P.O.  BOX  68,  FIN-80101 
Joensuu,  Finland 


631 


of  each  sample  tree  characteristic  with  respect  to  measured  characteristics  is 
given.  Most  methods  used  are  based  on  regression  techniques  (Korhonen  1993, 
Korhonen  1992,  Cunia  1986,  Kilkki  1979).  The  advantage  of  using  regression 
models  is  that  unbiased  estimates  for  the  population  parameters  are  easily 
obtained. 

The  unbiased  estimate  of  mean  does  not  suffice,  for  example,  if  the  predicted 
values  are  used  as  a  data  base  for  simulating  the  future  development  of  forests 
(Ranneby  and  Svensson  1990).  It  is  especially  important  to  retain  the  natural 
variation,  if  the  models  used  for  predicting  future  development  are  non-linear 
(Moeur  and  Stage  1995).  In  this  case,  the  different  tree  characteristics  should 
harmonize,  the  between-plot  and  within-plot  variation  of  the  predictions  should  be 
as  realistic  as  possible,  the  treewise  and  standwise  results  should  be  unbiased  and, 
in  addition,  the  mean  in  the  population  should  be  unbiased. 

These  requirements  are  often  contradictory.  It  is  difficult  to  formulate  the 
model  so  that  it  is  precise  enough,  provides  realistic  image  of  forests  and  gives 
unbiased  estimates  for  arbitrary  groups.  A  method,  which  is  optimal  in  one 
respect  may  be  unacceptable  in  other  respects.  What  is  acceptable,  depends  on  the 
situation. 

One  example  of  a  problem  of  this  kind  is  describing  the  effect  of  geographic 
location  on  stem  form.  In  the  paper  of  Korhonen  (1993)  it  was  demonstrated  that, 
in  Finland,  the  stem  form  of  Scots  pine  (Pinus  sylvestris)  depends  on  the 
geographic  location  besides  the  tree  and  stand  characteristics.  The  traditional 
solution  for  this  problem  is  to  estimate  separate  models  for  areas  of  interest.  In 
this  approach,  however,  the  effect  of  location  cannot  be  taken  into  account  for 
areas  smaller  or  larger  than  the  areas  pre-defined  and  the  effect  of  location  is  not 
continuous.  Another  solution  is  to  take  the  location  as  a  regressor  to  the  applied 
model.  A  quadratic  trend  surface,  for  example,  can  be  used  for  describing  the 
large  scale  variation  (Korhonen  1993).  With  the  kriging  method  also  the  small- 
scale  correlations  can  be  taken  into  account  (Ripley  1981).  Unfortunately,  the 
kriging  method  may  be  impractical  for  large  data  sets. 

Often  the  models  for  generalizing  different  sample  tree  characteristics  are 
estimated  separately.  Thus,  it  is  possible  to  obtain  values  for  volume  and  volume 
increment,  which  result  in  unrealistic  estimate  for  volume  increment  percentage. 
Standard  estimation  procedures  minimize  componentwise  losses,  but  what  is 
needed  is  an  ensemble  of  parameters  (Louis  1984).  Solution  for  this  situation  may 
be  a  system  of  simultaneous  equations. 

Non-parametric  models  can  offer  flexible  solutions  for  generalizing  sample  tree 
data  for  different  kind  of  situations.  With  non-parametric  methods  the  effect  of 
location  can  be  taken  into  account  with  a  simple  procedure.  It  is  also  easy  to 
generalize  all  the  sample  tree  characteristics  at  the  same  time.  Thus,  it  is  easier  to 
retain  the  covariance  structure  between  the  different  tree  characteristics.  In  this 
paper  an  application  of  semiparametric  kernel  regression  and  nearest-neighbor 
regression  for  generalizing  sample  tree  information  are  discussed.  The  goal  of  this 


632 


paper  is  to  compare  the  presented  non-parametric  methods  from  theoretical  point 
of  view  and  also  with  an  empirical  example. 


METHODS 


Kernel  regression  approach 


A  non-parametric  estimate  of  a  variable  y  at  a  point  i  is  the  weighted  average 
of  the  measured  values  of  y.  The  weight  of  a  sample  point  depends  on  the 
differences  in  the  values  of  the  independent  variables  between  the  point  of  interest 
and  the  sample  points.  In  this  study  a  non-parametric  regression  model  (Nadaraya 
1964) 

 _  (1) 

J     JL  (x-x) 
1=1  h 


was  used,  where  y  is  the  dependent  variable,  Xj  (Xj)  is  a  vector  containing  the 
values  of  the  independent  variables  at  point  j  (i),  h  is  the  window-parameter  and 
K  is  the  kernel  function.  In  this  study  the  kernel  function  used  was  a  multivariate 
normal  density  function  (Silverman  1986) 


where  d  is  the  dimension  of  the  distribution  (number  of  elements  in  vector  Xj  in 
1). 

The  use  of  non-parametric  kernel  regression  in  the  case  of  several  independent 
variables  is,  however,  difficult.  When  the  number  of  independent  variables 
increases,  the  data  set  may  be  surprisingly  sparsely  distributed  in  a  high- 
dimensional  Euclidean  space.  Thus,  the  applicable  window-parameter  values  are 
too  large  to  describe  the  relationships  between  the  dependent  and  independent 
variables  properly.  Further,  the  model  becomes  difficult  to  interpret  and 
impossible  to  demonstrate  in  a  graphic  form.  (Hardle  1989). 

Several  methods  for  overcoming  this  problem  have  been  presented,  for 
example,  by  considering  the  linear  combinations  of  the  independent  variables  (see 
Hardle  1989,  Moeur  and  Stage  1995).  One  obvious  solution  for  the  problem  of 
several  independent  variables  is  to  use  semiparametric  methods,  i.e.  combination 
of  parametric  and  non-parametric  methods.  In  this  study,  the  residuals  of  a 
parametric  model  were  smoothed  with  non-parametric  kernel  regression  with 
coordinates  as  independent  variables,  in  order  to  obtain  localized  estimates. 


633 


Nearest-neighbor  approach 


A  nearest  neighbor  estimator  is  a  (weighted)  average  of  k  nearest  neighbors  of 
point  j. 

E  yFbj*) 

**}=-k   <3> 

E 

i=i 

In  this  study  the  weight  function  used  was 

p  (4) 


1  + 


where  a  and  k  are  parameters.  The  nearest  neighbors  are  defined  by  some 
distance  measure,  in  this  study  the  nearest  neighbors  are  those  with  largest 
weights.  The  parameter  an,  defines  the  relative  importance  of  the  independent 
variable  xm.  Another  parameter  of  the  nearest  neighbor  estimator  (Eq.  3)  is  the 
number  of  neighbors  included,  k. 

The  kernel  method  and  nearest-neighbor  method  are  closely  related  methods. 
In  nearest-neighbor  approach  the  size  of  the  neighborhood  may  vary,  whereas  in 
non-parametric  kernel  regression  the  size  of  the  neighborhood  is  fixed  and  the 
number  of  neighbors  varies.  Nearest  neighbor  method  is  equivalent  to  kernel 
method  with  varying  window  width.  The  dimensionality  problem  in  the  nearest 
neighbor  approach  does  not  seem  to  be  as  serious  problem  as  in  the  kernel 
method,  because  of  the  varying  window  width. 

The  methods  differ  also  with  other  respects.  For  example,  the  nearest-neighbor 
produces  a  slightly  rougher  curve  (Hardle  1989).  This  is  due  to  discontinuity  of 
the  function.  As  the  window  moves,  new  observations  enter  to  the  k  neighbors. 
If  the  new  neighbor  differs  from  the  current  average,  there  will  be  an  abrupt 
change  in  the  value  of  nearest-neighbor  regression.  Also,  due  to  the  fixed  number 
of  neighbors  nearest  neighbor  method  can  be  expected  to  be  more  biased  near  the 
boundaries  than  the  kernel  method. 

Optimal  values  of  the  parameters  can  be  searched  with  cross  validation 
method  (Altman  1992).  In  this  method  each  observation  is  predicted  with  data 
excluding  the  observation  itself.  Estimator  with  lowest  mean  of  squared  residuals 
is  regarded  best.  In  this  study,  the  parameters  used  in  this  study  are  from 
Korhonen  and  Kangas  (1995). 


634 


AN  EXAMPLE 


Material 

The  data  used  in  this  study  were  the  pine  sample  trees  of  8th  NFI  of  eastern 
Finland.  Only  trees  growing  on  site  class  II  were  included  in  the  data.  The  data 
consist  of  2063  pines  measured  on  375  plots. 

Diameter  at  breast  height  and  height  were  measured  for  the  sample  trees  used 
in  this  study.  Stem  volumes  were  calculated  using  measured  dimensions  and 
volume  functions  of  Laasasenaho  (1982).  For  each  plot  several  variables 
describing  the  site  and  growing  stock  were  registered  in  the  NFI  data.  These 
variables  include  location,  altitude,  site  class,  basal  area  of  growing  stock, 
dominant  tree  species,  mean  diameter  and  age  of  growing  stock  etc. 

Results 

The  nearest  neighbor  estimator  was  compared  with  the  parametric  and 
semiparametric  estimator.  The  volumes  of  sample  trees  of  NFI8  were  estimated 
with  other  trees  in  the  same  data  set.  Mean  and  standard  deviation  of  residuals  of 
volume  estimates  were  calculated  using  parametric  estimator  (Eq.  5),  nearest 
neighbor  estimator  (Eqs.  3,  4)  and  semiparametric  estimator  (Eqs.  1,  2)  with 
different  values  of  parameters.  The  within-plot  and  between-plot  variance 
components  of  volume  were  estimated  to  test  how  the  different  methods  retain  the 
initial  variation  in  the  data. 

As  the  parametric  estimator  and  the  parametric  part  of  the  semiparametric 
model  a  function 

was  used.  In  Eq.  5  v«  is  the  volume  of  tree  i  on  plot  j  (dm3),  is  the  diameter  at 
breast  height  (cm),  ln(Gj)  is  the  natural  logarithm  of  the  basal  area  of  the  growing 
stock  of  the  plot  j  (m2/ha),  Tj  is  the  mean  age  and  d^  is  the  mean  diameter  in  the 
plot  j.  The  only  independent  variables  in  the  non-parametric  part  of  the  model 
were  the  coordinates.  For  the  nearest  neighbor  approach  the  independent  variables 
were  d2^,  GJ9  Tj5  d^  and  the  coordinates.  The  b-parameters  for  parametric  model 
and  weights      for  nearest  neighbor  approach  are  presented  in  table  1. 


635 


Table  1.  The  variables  in  parametric  model  and  their  coefficients  and,  the  variables  in  the 
nearest-neighbor  regression  and  their  weights. 


Variable 

Coefficient 

Variable 

Weight 

int 

-0.0977 

d2 

0.227 

d 

0.0274 

G 

2.020 

d2 

-0.00035 

T 

1.812 

ln(G) 

0.0576 

1.937 

T 

0.00102 

X 

0.149 

0.00792 

y 

0.053 

The  results  were  calculated  with  several  different  window  widths  h  for 
semiparametric  approach  and  with  different  number  of  neighbors  k  for  nearest 
neighbor  approach.  The  results  are  presented  in  tables  2  and  3. 

Table  2.  Mean  and  standard  error  of  residuals  of  volume  and  between-plot  and  within-plot 
standard  deviations  of  volume  predictions  with  different  window  widths  for  semiparametric 
approach.  The  between-plot  and  within-plot  standard  deviations  of  true  volumes  are  254.1 
and  179.6,  respectively. 


Window 

Mean  of 

Std.  error  Std(plot) 

Std(tree) 

width 

residuals 

2  km 

-2.19 

49.78  249.3 

181.5 

4  km 

-1.52 

52.66  248.6 

181.3 

8  km 

-0.99 

57.67  248.8 

181.3 

16  km 

-1.02 

63.36  250.8 

181.1 

parametric 

0.0 

66.36  248.5 

181.5 

Table  3.  Mean  and  standard  error  of  residuals  of  volume  and  between-plot  and  within-plot 
standard  deviations  of  volume  predictions  with  different  number  of  neighbors  for  nearest- 
neighbor  approach.  The  between-plot  and  within-plot  standard  deviations  of  true  volumes 
are  254.1  and  179.6,  respectively. 


Number  of 
neighbors 

Mean  of 
residuals 

Std.  error 

Std(plot) 

Std(tree) 

1 

0.26 

77.38 

252.6 

175.9 

5 

-0.54 

65.96 

248.1 

171.2 

10 

0.058 

65.40 

246.7 

170.2 

15 

0.52 

65.84 

245.1 

169.4 

The  parametric  estimates  of  volume  are  unbiased  in  the  whole  area,  whereas 


636 


the  non-parametric  estimates  generally  are  not  unbiased.  The  smallest  standard 
errors  are  obtained  with  semiparametric  approach,  using  quite  small  window 
widths.  The  largest  standard  errors  were  obtained  with  nearest-neighbor  method. 

On  the  other  hand,  with  parametric  regression  estimator  the  between-plot 
variation  decreases  and  within-plot  variation  increases,  when  compared  to  the  true 
variations.  The  semiparametric  approach  has  no  effect  at  all  on  the  variance 
components.  The  most  realistic  variation  is  obtained  with  nearest-neighbor  method 
with  only  one  neighbor.  From  these  results  it  can  be  concluded  that  the  preferable 
method  for  generalizing  sample  tree  information  depends  on  the  situation. 

DISCUSSION 

In  the  standard  estimation  methods,  like  regression  analysis,  the  primary  goal 
is  to  obtain  as  accurate  estimates  as  possible  for  individual  observations.  Thus, 
variance  or  MSE  of  individual  predictions  is  minimized.  In  forest  inventory, 
however,  the  goal  is  to  obtain  unbiased  estimates  of  several  parameters  for 
arbitrary  groups  of  observations.  Consequently,  minimizing  the  variance  or  MSE 
is  not  enough.  The  other  criterias  may  be  even  more  important  than  the  MSE  of 
individual  observations.  Also,  the  different  requirements  that  are  set  to  the  results 
of  generalization  method  may  be  contradictory. 

In  this  study,  in  addition  to  the  variance  and  bias  also  within-plot  and  between- 
plot  variation  of  the  predictions  were  considered.  Other  components,  such  as  sub- 
group biases  and  variances  could  also  have  been  considered.  In  this  study  the 
different  components  were  not  combined  in  order  to  find  an  optimal  method,  but 
it  would  be  possible  to  define  a  "utility  function"  of  the  different  components  and 
their  relative  weights. 

With  the  semiparametric  approach  the  estimates  of  sample  tree  characteristics 
can  be  localized  in  order  to  obtain  better  estimates  for  subareas  (-groups)  of  data 
(Kangas  and  Korhonen  1995).  This  is  also  true  for  nearest  neighbor  method.  But 
in  addition,  with  the  nearest  neighbor  method  the  initial  structure  of  the  data  can 
be  retained  fairly  well  (see  also  Korhonen  and  Kangas  1995).  This  is  due  to  the 
fact  that  all  the  sample  tree  characteristics  can  be  generalized  at  the  same  time. 
With  a  purely  non-parametric  kernel  regression  method  this  would  also  be 
possible.  This  approach,  however,  requires  special  attention  to  dimensionality 
problem,  for  example  smoothing  in  one  dimension  with  a  linear  combination  of 
variables. 

REFERENCES 

Altman,  N.S.  1992.  An  Introduction  to  Kernel  and  Nearest-neighbor 
Nonparametric  Regression.  The  American  Statistician  46:175-185. 

Cunia,  T.  1986.  Error  of  forest  inventory  estimates:  its  main  components.  In: 
Estimating  Tree  Biomass  Regressions  and  Their  Error.  Proceedings  of  the 


637 


Workshop  on  Tree  Biomass  Regression  Functions  and  their  Contribution  to  the 
Error  of  Forest  Inventory  Estimates.  May  26-30,  1986.  Syracuse,  New  York, 
pp.  1-13. 

Nadaraya,  E.A.  1964.  On  Estimating  Regression.  Theory  of  Probability 

Application.  No.  9:141-142. 
Handle,  W.  1989.  Applied  nonparametric  regression.  Cambridge  University  Press. 

323  pp. 

Kangas,  A.  and  Korhonen,  K.T.  1995.  Generalizing  sample  tree  information  with 

semiparametric  and  parametric  models.  Silva  fennica  29(2):  151-158. 
Kilkki,  P.  1979.  An  outline  for  a  data  processing  system  in  forest  mensuration. 

Silva  Fennica  13(4):368-384. 
Korhonen,  K.T.  1992.  Calibration  of  upper  diameter  models  in  large  scale  forest 

inventory.  Silva  Fennica  26(4):23 1-239. 
Korhonen,  K.T.  1993.  Mixed  estimation  in  calibration  of  volume  functions  of 

Scots  pine.  Silva  Fennica  27(4):269-276. 
Korhonen,  K.T.  and  Kangas,  A.  1995.  Application  of  nearest-neighbor  regression 

for  generalizing  sample  tree  information.  Manuscript,  16  p. 
Laasasenaho,  J.  1982.  Taper  curve  and  volume  functions  for  pine,  spruce  and 

birch.  Communicationes  Instituti  Forestalis  Fenniae  108.  74  pp. 
Louis,  T.A.  1984.  Estimating  a  population  of  parameter  values  using  Bayes  and 

empirical  Bayes  methods.  J.  Am.  Stat.  Ass.  79(386):393-398. 
Moeur,  M.  and  Stage,  A.R.  1995.  A  most  similar  neighbor:  an  improved  sampling 

inference  procedure  for  natural  resource  planning.  For.  Sci  41(2):337-359. 
Ranneby,  B.  and  Svensson,  S.A.  1991.  From  sample  tree  data  to  images  of  tree 

populations.  In:  Forest  inventories  in  Europe  with  special  reference  to  statistical 

methods.  Proceedings  of  the  International  IUFRO  S.  4.02  and  S.  604 

Symposium,  May  14-16,  1990.  Swiss  Federal  Institute  for  Forest,  Snow  and 

Landscape  Research.  Birmensdorf,  Switzerland. 
Silverman,  B.W.  1986.  Density  Estimation  for  Statistics  and  Data  Analysis. 

London.  Chapman  &  Hall. 

BIOGRAPHICAL  SKETCH 

Annika  S.  Kangas  is  research  scientist,  Finnish  Forest  Research  Institute, 
Kannus  Research  Station,  Finland.  She  holds  a  D.Sc.  in  Forestry  from  University 
of  Joensuu.  Annika  deals  with  forest  inventory  methods,  especially  with  model 
based  sampling  techniques. 

Kari  T.  Korhonen  is  research  scientist,  Finnish  Forest  Research  Institute, 
Joensuu  Research  Station,  Finland.  He  holds  a  D.Sc.  in  Forestry  from  University 
of  Joensuu.  Kari  deals  with  forest  inventory  methods,  especially  with  generalizing 
sample  tree  information. 


638 


A  Method  for  Assessing  the  Prediction 
Quality  of  Mechanistic  Forest 
Growth  Models 

Biing  T.  Guan1,  George  Gertner2  and  Pablo  Parysow3 


Abstract.  A  method  for  assessing  the  prediction  quality  of  mechanistic 
forest  growth  models  was  presented.  The  method  consisted  of  four  steps: 
assuming  distributions  for  parameter  values,  parameter  screening, 
outlining  model  behavior  through  sampling,  and  approximating  model 
behavior  based  on  the  sampled  points.  The  proposed  method  was  then 
applied  to  a  carbon  balance  stand  level  forest  growth  model.  For  the 
example,  a  Monte  Carlo  method  was  used  to  perform  the  sampling,  and 
then  a  "patterned"  artificial  neural  network  was  used  for  the 
approximation.  It  was  found  that  the  predictions  of  the  model  were 
relatively  unbiased  with  respect  to  the  initial  parameter  variances,  and 
the  variances  of  the  predictions  were  mainly  contributed  by  only  a  few 
parameters.  Such  information  allows  the  model  to  be  more  closely 
examined  and  provides  a  way  to  improve  the  model. 


INTRODUCTION 

For  the  past  ten  years,  an  important  change  in  forest  growth  modeling  has  been 
the  shift  in  modeling  approach  from  empirical  modeling  towards  mechanistic  (or 
process)  modeling;  a  shift  that  reflects  the  model  developers '  desire  to  try  to  gain 
better  insights  into  how  stands  (or  trees)  grow  and  how  future  environment 
changes  may  affect  forest  ecosystems  (Bossel,  1991).  Though  still  not  having  the 
desired  predictive  power,  mechanistic  forest  growth  models  are  being  developed 
and  used  more  frequently  to  answer  some  serious  questions  regarding  the  likely 
impact  of  global  environmental  changes  on  various  aspects  of  forest  ecosystems. 
An  example  is  the  use  of  TREGRO,  a  mechanistic  tree  growth  model,  to  assess 
the  impact  of  03  on  seedling  growth  of  red  spruce  (Laurence  et  ai,  1993).  Such  a 
usage  of  mechanistic  models  raises  an  important  question:  what  is  the  prediction 
quality  (i.e.,  the  amount  of  prediction  bias  and  variance)  of  a  mechanistic  model 


Associate  Professor,  Department  of  Forestry,  National  Taiwan  University,  Taipei,  Taiwan,  Republic  of  China 
'  Professor,  Department  of  Natural  Resources  and  Environmental  Sciences,  University  of  Illinois  at  Urbana-Champaign 
Urbana.  IL 

Ph.D.  graduate  student,  Department  of  Natural  Resources  and  Environmental  Sciences,  University  of  Illinois  at 
Urbana-Champaign,  Urbana,  IL 


639 


and  how  can  such  information  be  obtained?  Existing  prediction  quality  analysis 
methods,  such  as  a  Monte  Carlo  approach  or  an  error  propagation  approach  (e.g., 
Gertner,  1987),  cannot  handle  mechanistic  models  effectively. 

In  this  study,  a  method  is  proposed  for  assessing  the  prediction  quality  of 
mechanistic  growth  models.  The  framework  of  the  method  and  layout  of  its  key 
steps  will  be  described  first.  Then,  as  an  example,  the  method  is  applied  to  a 
mechanistic  growth  model.  Finally,  some  comments  will  be  made  with  respect  to 
the  generality  of  the  proposed  method. 

GENERAL  FRAMEWORK  OF  THE  METHOD 

The  core  idea  of  the  proposed  prediction  quality  analysis  method  is  as  follows: 
we  believe  that  the  sampling  intensity  required  by  a  Monte  Carlo  approach  can  be 
reduced  through  the  use  of  a  good  approximation  procedure  while  maintaining  a 
specified  accuracy  of  the  quality  analysis.  We  should  emphasize  at  this  point  that 
the  proposed  method  assumes  that  the  target  model  is  structurally  and  behaviorally 
valid  (see  Bossel,  1992)  for  the  intended  purposes.  The  proposed  method  has  the 
following  four  key  steps: 

Assuming  Ranges  and  Distributions:  For  the  proposed  method,  we  assume  each 
parameter  in  the  model  has  a  fixed  range  and  a  distribution  within  the  specified 
range.  Unlike  empirical  models,  parameters  in  mechanistic  models  are  usually  not 
estimated  from  data  (some  even  argue  that  parameters  in  mechanistic  models 
should  not  be  estimated,  e.g.,  Bossel,  1992);  they  are  usually  derived  from  the 
literature,  small  scale  experiments,  studies  of  related  species,  and  from  subjective 
sources  when  information  is  sparse.  Though  often  viewed  as  exact,  parameter 
values  in  mechanistic  models  should  in  reality  be  considered  as  interval  estimates, 
that  is,  the  "true"  parameter  value  may  be  within  a  certain  range,  and  the  given 
parameter  value  is  the  most  likely  a  value  within  the  range. 

Screening  Parameters:  Parameter  screening  is  necessary  for  reducing  the  actual 
number  of  parameters  that  need  to  be  examined.  It  is  likely  that  in  a  mechanistic 
model  not  all  parameters  contribute  equally  to  the  prediction  quality  (bias  and 
variance)  of  that  model.  Some  parameters  may  have  significant  effects  on 
prediction  quality,  some  may  have  slight  effects,  and  still  others  may  have  no 
effects.  If  those  parameters  that  contribute  little  to  prediction  quality  are  screened 
out,  then  computational  resources  can  be  saved  in  the  subsequent  steps. 

Determining  Monte  Carlo  Sampling  Intensity:  For  the  proposed  method,  two 
sampling  intensities  need  to  be  considered.  The  first  intensity  is  the  total  number 
of  points  (each  point  represents  a  combination  of  variabilities  of  the  parameters 


640 


and  the  initial  values  of  the  state  variables)  that  will  be  used  in  the  subsequent 
approximation.  The  sampled  points  roughly  define  the  model '  s  global  behavior  in 
the  sampled  space.  The  second  intensity  is  the  number  of  runs  that  need  to  be 
carried  out  at  each  sampled  point  to  obtain  actual  prediction  quality  information  at 
that  sampled  point  (i.e.,  the  local  behavior  around  the  sampled  point). 

Selecting  an  Approximation  Procedure:  Together  with  the  previous  step,  this 
step  determines  the  success  of  the  proposed  method.  Choices  abound  for  selecting 
an  approximation  procedure.  One  can  choose  from  simple  polynomials  to 
complex  nonlinear  procedures,  such  as  a  spline  or  a  neural  network  model. 
Different  approximation  procedures  offer  different  advantages. 

EXAMPLE 

The  model  that  we  used  to  demonstrate  the  proposed  method  is  a  mechanistic 
forest  growth  model  developed  by  Valentine  (1988).  The  model  is  a  carbon 
balance  model  based  on  the  pipe  theory  (Shinozaki  et  al.,  1964  a.b)  and  the  self- 
thinning  rule.  The  model  is  a  stand-level  growth  model  for  even-aged  stands.  It 
has  three  state  variables  (basal  area  per  unit  area,  active  pipe  length  per  unit  area, 
and  total  woody  volume  per  unit  area)  and  20  parameters.  In  this  example,  the 
pipe  model  was  adapted  for  red  pine  {Pinus  resinosa  Ait.).  The  model  was  used  to 
simulate  the  growth  of  red  pine  for  30  years,  from  age  30  to  age  60.  For  this 
example,  we  will  view  the  means  as  the  "true"  parameter  values.  Therefore,  the 
final  projections  based  on  those  means  will  be  treated  as  the  "unbiased" 
projections.  Bias  will  thus  be  defined  as  the  differences  between  the  projected  and 
the  "unbiased"  values. 

Step  1:  For  the  example,  the  mean  value  and  the  lower  and  the  upper  bounds  for 
each  of  the  parameters,  as  well  as  for  the  initial  values  of  the  state  variables,  were 
directly  taken  from  Parysow  (1994).  It  is  assumed  that  each  parameter,  as  well  as 
the  initial  value  of  a  state  variable,  has  an  unimodal  and  symmetric  beta 
distribution  bounded  between  the  prescribed  lower  and  upper  bounds.  The 
requirements  on  the  shape  of  our  beta  distributions  constrains  the  maximum 
variance  that  our  beta  distributions  can  have.  For  a  beta  distribution  to  be 
symmetric  and  unimodal.  its  two  shape  parameters  need  to  be  equal,  and  the 
values  must  be  greater  than  1.  In  this  example,  it  was  decided  that  the  minimum 
value  for  the  shape  parameters  of  a  beta  distribution  would  be  1.5  which  gives  a 
beta  distribution  with  a  well-defined  mode.  Under  these  constraints,  the  standard 
deviation  of  a  generic  beta  distribution  in  our  analysis  cannot  be  greater  than  0.25; 
in  other  words,  the  maximum  standard  deviation  of  a  parameter  distribution  is  25 
percent  of  its  distribution  range. 


641 


Step  2:  Though  the  target  model  was  a  small  model  in  terms  of  the  number  of 
parameters  and  state  variables,  it  was  still  desirable  to  screen  out  parameters  that 
had  little  influence  on  projection  quality.  The  procedure  used  is  Mann '  s  test  for  a 
trend  (Lehmann,  1975,  pp.  290-297).  Parameters  were  screened  independently, 
and  the  state  variables  were  not  included  in  the  screening  procedure.  The  a-level 
to  include  was  set  to  be  0.2  since  the  intention  was  to  screen  out  the  parameters 
with  little  influences  on  the  final  projection  quality. 

Step  3:  In  this  example,  we  chose  a  random  sampling  scheme  to  generate  the 
necessary  sample  points  as  well  as  the  local  behavior  at  each  sampled  point.  Two 
sets  of  data  were  generated.  For  the  first  set,  the  calibration  data  set,  the  sampling 
intensity  was  64K  for  the  first  level  sampling  and  16K  for  the  second  level 
sampling.  For  the  second  set,  the  validation  data  set,  the  sampling  intensities  were 
128K  and  32K,  respectively,  for  the  first  and  second  levels  of  sampling.  It  should 
be  noted  that  although  parameter  screening  was  performed  (Step  2),  a  complete 
sample  was  still  done  because  the  approximation  results  were  used  to  validate  the 
screening  results. 

Step  4:  The  approximation  procedure  used  in  this  example  was  a  feedforward, 
multilayer  type  of  artificial  neural  network  model.  Artificial  neural  networks  have 
been  proven  both  theoretically  (Cybenko,1989;  Hornik  et  ai,  1989)  and 
empirically  (e.g.,  Lapedes  and  Farber,  1987)  to  be  excellent  function 
approximators  under  general  conditions.  Artificial  neural  networks  have  also  been 
used  to  approximate  chaotic  functions  with  good  results  (e.g.,  Lapedes  and  Farber, 
1987).  Though  chaotic  behavior  might  not  exist  in  this  example,  it  may  exist  in 
other  mechanistic  models,  especially  the  complex  ones,  since  most  of  the 
differential  equations  in  existing  mechanistic  models  have  not  been  examined 
thoroughly  for  their  behaviors.  For  these  reasons,  artificial  neural  networks  have 
been  chosen  as  the  approximation  tools  in  this  example. 

The  networks  used  in  this  example  all  had  an  input  layer,  an  output  layer  and  a 
hidden  layer.  The  networks  used  were  different  from  the  typical  feedforward 
networks  in  two  aspects.  First,  the  nodes  between  the  input  layer  and  the  hidden 
layer  were  not  fully  connected,  and  were  without  the  bias  connections  (Figure  1). 

The  nodes  in  the  hidden  layer  were  divided  into  groups,  and  each  group  was 
connected  only  to  one  input  node  (groups  were  disjoint).  Thus,  each  input  was 
trained  by  a  subset  of  hidden  nodes.  The  activation  function  for  the  hidden  nodes 
was  the  arc  tangent  (tan-1)  function.  The  output  activation  function  was  just  a 
linear  function  (i.e.,  no  squashing). 


642 


Figure  1.  Network  architecture  for  the  proposed  network. 

Such  a  network  architecture  will  produce  outputs  that  can  be  partitioned 
according  to  the  contribution  of  each  input.  The  connection  pattern  of  the  network 
will  isolate  the  contribution  of  each  input,  while  the  arc  tangent  activation 
function  will  make  some  hidden  nodes  inactive  (i.e.,  set  to  O's)  when  the 
corresponding  inputs  are  inactive.  The  contributions  from  those  inactive  inputs  to 
the  overall  network  outputs  will  thus  be  O's.  To  be  able  to  partition  the  outputs 
according  to  the  contribution  of  each  input  is  important  to  the  proposed  procedure 
since  it  allows  us  to  examine  the  importance  of  each  input  and  to  develop  an  error 
budget  for  the  model. 

The  network  we  used  in  this  example  had  five  hidden  nodes  for  each  input 
node.  The  training  algorithm  used  in  this  case  study  was  a  modified  version  of  a 
random  optimization  procedure  (Baba,  1989).  The  error  function  of  the  training 
method  was  the  squared  error  function  (i.e.,  the  squared  difference  between  the 
target  and  the  actual  outputs).  Since  this  particular  example  involves  a  large 
number  of  training  and  validation  data,  and  the  selected  training  algorithm  is 
parallel  in  nature,  network  training  as  well  as  validation  were  conducted  on  a 


643 


CM-2  parallel  computer.  Detail  discussion  of  the  training  algorithm  can  be  found 
in  Gumetal.  (1993). 

RESULTS  AND  COMMENTS 

Since  the  results  from  the  sampling  showed  that  initial  parameter  variability  did 
not  affect  the  projection  accuracy  (i.e.,  very  little  bias)  of  the  pipe  model,  we  will 
focus  on  how  initial  parameter  uncertainty  affects  the  projection  variances  of  the 
pipe  model. 

Parameter  Screening:  The  results  from  parameter  screening  confirm  our  initial 
hypothesis  that  not  all  the  parameters  contribute  significantly  toward  the 
projection  variance  of  the  pipe  model.  Table  1  lists  the  parameters  that  the  initial 
screening  procedure  deemed  to  be  significant  (with  p  value,  or  calculated  tail 
probability,  smaller  than  0.2)  to  overall  projection  variance  for  each  of  the  three 
state  variables.  The  entries  are  listed  in  a  descending  order  according  to  the 
calculated  p  value  of  each  parameter. 

Approximation  Results:  Table  2  gives  the  validation  results  of  the  trained  neural 
network  for  approximating  the  final  projection  variances  of  the  three  state 
variables.  The  mean  error  rates  are  within  4  percent  for  all  the  three  state 
variables,  and  99  percent  of  the  approximations  are  within  14  percent  of  the  target 
values.  It  should  be  noted  that  the  trained  network  was  for  approximating  the 
variances  of  the  three  state  variables  simultaneously  (i.e.,  the  network  had  three 
output  nodes). 

Ranking  the  Importance  of  Parameters:  Since  the  neural  network  model  we 
developed  has  the  ability  to  isolate  the  variance  contribution  of  an  individual 
parameter,  we  can  use  this  feature  to  rank  the  importance  of  individual 
parameters.  Table  3  is  an  example  of  such  a  ranking.  In  the  table,  we  can  see  that 
given  the  same  amount  of  relative  initial  variability,  the  one  with  the  largest 
contribution  is  the  parameter  S,  a  parameter  that  represents  the  effects  of 
environmental  conditions  on  tree  growth.  What  is  interesting  is  that  the  first  four 
parameters  in  the  table  account  for  almost  75  percent  of  the  total  variability,  and 
the  first  ten  parameters  account  for  about  95  percent  of  the  total  projection 
variance.  The  partitions  of  projection  variances  of  the  other  two  state  variables 
also  show  similar  patterns,  though  with  different  parameter  ranking.  Common  to 
the  projection  variances  of  the  three  state  variables  is  the  relative  importance  of 
the  parameter  S.  Given  the  same  amount  of  relative  initial  parameter  variability, 
this  parameter  accounts  for  roughly  37  percent,  63  percent  and  55  percent  of  the 
final  projection  variances  for  the  state  variables  basal  area,  pipe  length  and  total 


644 


Table  1.  Results  of  parameter  screening  for  the  final  projection  variance  for  the  three  state 
variables.  Entries  are  listed  in  a  descending  order  according  to  the  calculated  p  value  of  each 

parameter.  

State  Variable 


Pipe  Length  Basal  Area  Total  Volume 

s  s  s 

z  c  z 

Bp  Amax  Amax 

P  Z  P 

e  bp  bp 

Tf  P  C 

x  e  Tf 

Tz  Tf  VI 

Pf  y  Tz 

a  X  9 

Amax  F  \|/ 

Bf  Tz  F 

a 


Table  2.  Validation  results  for  the  trained  neural  network  for  predicting  the  final  projection 
variance  of  the  three  state  variables.  Numbers  represents  the  relative  error  rates  (in 

percentage)  

State  Variables 


Pipe  Length  Basal  Area  Total  Volume 

Mean  3.79  1.38  2.03 

Minimum  0.00  0.00  0.00 

25Quantile  1.70  0.52  0.76 

Medium  3.38  1.12  1.61 

75Quantile  5.24  1.96  2.82 

95Quantile  8.79  3.52  5.36 

99Quantile  13.05  4.92  7.83 

Maximum  30.04  11.68  17.23 


volume,  respectively.  Such  a  ranking  in  parameter  contribution  provides  us  with 
clues  to  further  improve  the  performance  and  projection  quality  of  a  model. 

As  the  contribution  of  an  individual  parameter  toward  the  overall  prediction 
quality  becomes  available,  we  can  use  such  information  to  develop  error  budgets 
for  the  model,  to  test  hypotheses,  and  to  analyze  the  statistical  power  of  the  tests, 
steps  that  are  all  necessary  in  using  mechanistic  models  to  make  informed 
decisions. 


645 


Table  3.  An  example  of  a  partition  of  the  final  projection  variance  of  total  basal  area  per  unit 
area  based  on  the  contribution  of  each  parameter.  The  initial  standard  deviation  for  each 
parameter  was  set  at  5  %  of  the  parameter '  s  range. 


Parameter 

Contribution 

Relative 

IIIIIJUI  LctHLc  \  /C  ) 

Cumulative 
imporidnic  \  /c ) 

S 

2.70 

37.52 

37.52 

c 

1.30 

18.07 

55.59 

Basal  Area 

0.98 

13.54 

69.14 

Z 

0.48 

6.63 

75.77 

Amax 

0.42 

5.82 

81.59 

Pipe  Length 

0.29 

3.98 

85.57 

Bp 

0.25 

3.46 

89.03 

P 

0.18 

2.52 

91.54 

0.11 

1.57 

93.11 

e 

0.09 

1.24 

94.35 

Total  Volume 

0.09 

1.22 

95.57 

Tf 

0.07 

1.04 

96.61 

T 

0.06 

0.85 

97.45 

V2 

0.06 

0.83 

98.29 

F 

0.03 

0.42 

98.71 

Tz 

0.02 

0.35 

99.05 

Rz 

0.02 

0.30 

99.35 

oc 

0.02 

0.27 

99  62 

Rp 

0.00 

0.14 

99.76 

Bf 

0.00 

0.12 

99.88 

Rf 

0.00 

0.08 

99.97 

VI 

0.00 

0.02 

99.99 

Bz 

0.00 

0.01 

100.00 

Total 

7.20 

100.00 

Comparing  the  rankings  in  Tables  1  and  3,  not  including  the  state  variables,  we 
can  see  that  the  parameter  screening  procedure  we  proposed  can  indeed  screen  out 
the  parameters  that  have  little  influence  on  the  final  projection  variance  of  total 
basal  area.The  proposed  screening  procedure  also  was  effective  in  screening  out 
the  parameters  that  had  little  effect  with  respect  to  the  other  two  state  variables. 
With  a  conservative  screening  criterion,  we  believe  that  the  proposed  screening 
procedure  can  effectively  reduce  the  number  of  parameters  that  need  to  be 
considered  while  still  maintaining  the  overall  approximation  accuracy.  When  the 


646 


model  of  interest  is  large  and  complex,  parameter  screening  may  save  a  large 
amount  of  computation  resources  in  the  future. 

As  mentioned  earlier,  many  approximation  procedures  can  be  used  for  the 
proposed  method,  and  one  such  method  is  orthogonal  response  surface  analysis. 
For  the  same  pipe  model,  we  have  also  developed  a  second-order  orthogonal 
response  surface  model  based  on  a  fractional  factorial  design  (Gertner  et  al., 
1996).  The  orthogonal  response  model  and  the  neural  network  model  yielded 
almost  identical  results,  though  the  former  requires  less  computation  since  it  was 
based  on  a  fractional  factorial  design.  However,  when  the  model  of  interest  has  a 
large  number  of  parameters,  or  when  the  higher  order  approximation  is  called  for 
(e.g.,  the  approximation  is  highly  non-linear),  then  it  is  our  belief  that  a  neural 
network  approach  may  be  more  appropriate.  A  major  difference  between  the  two 
models  lies  in  the  way  the  total  projection  variance  (or  bias)  is  partitioned.  For  an 
orthogonal  model,  we  may  not  be  able  to  isolate  the  contribution  of  an  individual 
parameter  since  interaction  with  other  parameters  may  be  significant.  Our  neural 
network  model,  on  the  other  hand,  has  the  ability  to  isolate  the  individual 
parameter  contribution,  and  the  partition  is  exact  (though  not  necessary 
orthogonal). 

The  method  proposed  in  this  study  is  also  general  in  the  sense  that  the  method 
can  be  used  as  effectively  in  analyzing  large  and  complex  models  as  in  small  and 
simple  models,  though  complex  or  large  models  may  require  more  computational 
resources.  For  example,  for  complex  multi-component  models,  we  can  adopt  a 
bottom-up,  divide-and-conquer  approach,  that  is,  we  analyze  the  behavior  of 
individual  components,  and  then  analyze  the  behavior  of  components  toward  the 
overall  model.  Thus,  it  is  our  belief  that  the  proposed  method  will  contribute 
significantly  toward  a  more  rational  use  of  mechanistic  models  in  ecological 
research. 

ACKNOWLEDGMENTS 

The  authors  thanked  the  National  Center  for  Supercomputing  Applications, 
University  of  Illinois  at  Urbana-Champaign,  for  granting  the  necessary  computer 
time  on  the  CM-2  computer. 


647 


REFERENCES 


Bossel,  H.,  1991.  Modeling  forest  dynamics:  Moving  from  description  to 

explanation.  Forest  Ecology  and  Management  42:  129-143. 
Bossel,  H.,  1992.  Real-structure  process  description  as  the  basis  of  understanding 

ecosystems  and  their  development.  Ecological  Modelling  63:  261-276. 
Cybenko,  G.,  1989.  Approximation  by  superposition  of  a  sigmoidal  function. 

Mathematics  of  Controls,  Signals  and  Systems  2:  303-314. 
Gertner,  G.,  1987.  Approximating  precision  in  simulation  projections:  an  efficient 

alternative  to  Monte  Carlo  methods.  Forest  Science  33:  230-239. 
Gertner,  G.,  P.  Parysow  and  B.  T.  Guan.  1996.  Projection  variance  partitioning  of 

a  conceptual  forest  growth  model  with  orthogonal  polynomial.  To  appear  in 

Forest  Science. 

Guan,  B.  T.  and  G.  Z.  Gertner,  and  D.  Kowalski.   1993.  Modeling  training  site 

vegetation  coverage  probability  with  a  random  optimization  procedure:  an 

artificial  neural  network  approach.  In:  the  Proc.  of  the  Conference  of 

Applications  of  Artificial  Neural  Networks  IV,  Orlando,  Florida,  pp.  682-688. 
Hornik,  K.,  Stinchcombe,  M.,  and  White,  H.,  1989.  Multilayer  feedforward 

networks  are  universals  approximators.  Neural  Network  2:  359-366. 
Lapedes,  A.,  and  Farber,  R.,  1987.  Nonlinear  signal  processing  using  neural 

networks:  Prediction  and  system  modelling.  Technical  Report  LA-UR-87- 

2662,  Los  Alamos  National  Laboratory,  New  Mexico. 
Laurence,  J.  A.,  Kohut,  R.  J.,  and  Amundson,  R.  G.,  1993.  Use  of  TREGRO  to 

simulate  the  effects  of  ozone  on  the  growth  of  red  spruce  seedlings.  Forest 

Science  39(3):  453-464. 
Lehmann,  E.  L.,  1975.  Nonpar ametrics:  Statistical  Methods  Based  on  Ranks. 

Holden-Day,  San  Francisco.  457  p. 
Parysow,  P.  F,  1994.  A  procedure  to  approximate  prediction  uncertainty  in 

ecological  conceptual  models.  Unpublished  M.S.  Thesis,  University  of  Illinois 

at  Urbana-Champaign,  Urbana,  Illinois.  126  p. 
Shinozaki,  K.,  Yoda,  K.,  Hozumi,  K.  and  Kira,  T.,  1964a.  A  quantitative  analysis 

of  plant  form  -  the  pipe  model  theory.  I.  Basic  analysis.  Japanese  Journal  of 

Forest  Ecology  14:  97-105. 
Shinozaki,  K.,  Yoda,  K.,  Hozumi,  K.,  and  Kira,  T.,  1964b.  A  quantitative  analysis 

of  plant  form  -  the  pipe  model  theory.  II.  Further  evidence  of  the  theory  and  its 

application  in  forest  ecology.  Japanese  Journal  of  Forest  Ecology  14:  133-139. 
Valentine,  H.,  1988.  A  carbon  balance  model  of  stand  growth:  a  derivation 

employing  pipe-model  theory  and  the  self-thinning  rule.  Annals  of  Botany 

62:389-396. 


648 


Statistical  Analysis  of  Error 
Propagation  in  National  Level 
Carbon  Budgets 

C.  J.  Cieszewski1,  D.  P.  Turner2,  and  D.  L.  Phillips 


Abstract 

4  Existing  national- level  carbon  budget  studies  typically  do  not  con- 
duct formal  uncertainty  analyses.  In  cases  where  attempts  at  address- 
ing carbon  budget  uncertainties  have  been  made,  they  are  generally 
limited  to  informal  variations  of  sensitivity  studies.  The  reasons  for 
limited  effort  towards  uncertainty  analysis  in  carbon  budget  studies  in- 
clude: 1)  unknown  or  poor  statistical  properties  of  the  input  variables; 
2)  use  of  arbitrary  relationships  and  analytically  intractable  discrete 
relationships  in  carbon  budget  models;  3)  a  high  level  of  aggregation 
and  complexity  in  calculations;  and  at  times,  4)  the  interactive  nature 
of  the  developed  carbon  budget  models  which  does  not  lend  itself  to 
the  application  of  a  Monte  Carlo  approach.  Development  of  a  generic 
uncertainty  analysis  framework  for  carbon  budget  studies  is  proposed 
here.  The  key  component  will  be  a  flexible  error  propagation  program 
designed  for  operation  within  Monte  Carlo  simulations  and  using  statis- 
tical software.  The  simulation  software  will  have  an  open  architecture, 
such  that  it  can  be  implemented  at  a  user  defined  level  of  aggrega- 
tion with  regard  to  land  cover  type  in  a  variety  of  countries.  Error 
estimates  will  be  developed  for  variables  including  total  biomass  car- 
bon, net  ecosystem  production,  and  carbon  emissions  associated  with 
land  use  change.  The  error  propagation  program  will  include  a  library 
of  functions  and  difference  equations  useful  for  replacing  discrete  re- 
lationships such  as  growth  and  yield  tables,  and  providing  a  basis  for 
analytical  treatment  of  potential  error  propagation. 


forest  Biometrician,  ManTech  Environmental  Research  Services  Corp.,  Corvallis,  OR. 
2Forest  Ecologist,  ManTech  Environmental  Research  Services  Corp.,  Corvallis,  OR. 
3Research  Biologist,  U.S.  EPA  National  Health  &;  Environmental  Effects  Laboratory, 
Corvallis,  OR. 

4We  acknowledge  Sandra  Brown,  U.S.  EPA  National  Health  &  Environmental  Effects 
Laboratory,  Corvallis,  OR,  for  the  initiation  of  this  study. 


649 


INTRODUCTION 


The  Framework  Convention  for  Climate  Change  (FCCC)  calls  for  national- 
level  inventories  of  net  greenhouse  gas  emissions,  including  carbon  (CO2-C) 
sources  and  sinks  associated  with  forests  and  land-cover  change.  The  Inter- 
national Panel  on  Climate  Change  (IPCC)  has  subsequently  worked  with  the 
Organization  for  Economic  Cooperation  and  Development  (OECD)  to  produce 
an  initial  set  of  protocols  for  estimating  the  land  base  flux  (IPCC  1995).  A 
number  of  national-  level  studies  have  been  made,  e.g.,  Makundi  et  al  1992, 
Kauppi  et  al  (1992),  Kolchugina  and  Vinson  (1993),  Kurz  et  al  (1992,  1994), 
Maclaren  and  Wakelin  (1991),  Turner  et  al  (1993,  1995),  but  none  of  them 
conducts  formal  error  analyses  and  the  only  attempts  at  addressing  carbon 
budget  uncertainties  are  limited  to  informal  versions  of  sensitivity  analyses, 
e.g.,  Kurz  et  al  (1994). 

There  are  several  reasons  why  so  little  progress  has  been  made  in  the  issue 
of  uncertainty  analysis  in  the  carbon  budget  studies.  One  reason  is  the  limited 
information  about  properties  of  errors  in  input  variables  used  in  the  estimation 
of  greenhouse  gas  flux  (Robinson  1989).  It  is  impossible  to  estimate  statistical 
properties  of  propagated  errors  if  the  statistical  properties  of  the  input  errors 
are  not  known  and  a  formal  sensitivity  analysis  might  be  the  only  possibility 
to  properly  quantify  the  uncertainties  in  carbon  budget  models. 

Furthermore,  when  there  is  a  fair  amount  of  information  about  the  errors 
associated  with  the  input  variables,  their  statistical  properties  may  not  lend 
themselves  to  parametric  error  propagation  analysis.  For  example,  forestry 
data  are  usually  characterized  by  large  variation,  strong  correlation  between 
different  variables,  and  frequently  a  lack  of  normality  in  their  distributions. 
The  properties  of  distributions  of  errors  associated  with  such  variables  can 
be  analytically  intractable  within  even  moderately  complex  systems;  paramet- 
ric statistical  analysis  rely  on  normally  distributed  variables,  and  even  just 
a  correlation  between  analyzed  variables  dramatically  increases  the  relative 
complexity  of  the  involved  calculations. 

Another  problem  for  analytical  error  analysis  is  the  use  of  arbitrary  non- 
functional relationships  in  computations.  Such  systems  in  principle  are  ana- 
lytically intractable.  A  propagation  of  errors  can  be  analytically  traced  only 
within  continuous  mathematically  defined  systems  of  equations  on  a  limited 
level  of  complexity.  With  large  and  complex  systems  it  is  usually  more  prac- 
tical and  at  times  the  only  possibility  to  apply  simulation  methods  and  sensi- 
tivity analysis.  However,  practical  applications  of  simulation  methods  require 
that  models  be  written  in  forms  that  lend  themselves  to  large  numbers  of 
repetitions  in  the  model  executions  and  easily  changeable  input  selections. 

To  solve  these  problems  and  to  address  the  concerns  of  uncertainty  analy- 
sis on  the  carbon  budget  and  flux  computation  a  Monte  Carlo  simulation  ap- 
proach is  suggested.  This  paper  addresses  some  concerns  related  to  using  the 
Monte  Carlo  simulation  approach  and  describes  its  application  with  regards 
to  national  level  carbon  budgets  and  national  level  carbon  flux  estimation. 

TYPES  AND  SOURCES  OF  ERRORS 

The  types  of  errors  are  in  general  divided  into  absolute  errors  and  relative 
errors.  The  errors  can  be  systematic  or  random,  or  random  with  a  constant 


650 


bias.  In  the  carbon  budget  modeling  and  flux  estimation  the  errors  can  origi- 
nate from  the  data  collection,  i.e.,  inventory  and  data  compilation,  and  from 
the  carbon  budget  modeling  framework. 

There  are  three  major  sources  of  errors  (Cunia  1985,  Gertner  and  Kohl 
1992)  associated  with  the  inventory:  measurement  errors,  sampling  errors, 
and  regression  errors.  Sampling  errors  can  be  the  greatest  contribution  to 
total  error  contained  in  inventory  data.  Exact  estimation  of  inventory  errors 
is  difficult  since  the  actual  exact  quantities  measured  in  forestry  are  never  well 
known  and  definable.  Estimation  of  errors  for  inventory  data  can  be  based  on 
combination  of  analytical  methods  associated  with  sampling  procedures  and 
simulation  techniques  applied  to  data  compilation  calculations. 

Most  errors  generated/ propagated  within  carbon  budget  modeling  frame- 
work are  subject  to  such  sources  as: 

•  Errors  in  Given  Input  Data. 

•  Round-off  Errors  During  the  computations. 

•  Truncation  and  Discretization  Errors. 

•  Errors  Resulting  from  Simplifications  in  the  Mathematical  Model. 

•  "Human"  Errors  and  Machine  Errors. 

ASSUMPTIONS  ON  CARBON  BUDGET  DESIGNS 

Since  it  is  desirable  to  have  one  consistent  computational  schema  defining 
all  potentially  applicable  carbon  budgets,  the  framework  developed  here  will 
be  applicable  only  to  a  lowest  common  denominator  of  carbon  budget  ele- 
ments. The  lowest  common  denominator  is  assumed  to  be  the  lowest  available 
inventory  of  carbon  quantities.  This  means  that  even  in  the  most  information 
rich  carbon  budgets,  e.g.,  the  US  carbon  budget,  the  analysis  will  assume  avail- 
abilities of  error  distributions  of  carbon  quantities  entering  the  carbon  budget 
model  on  the  lowest  level  of  resolution,  e.g.,  live  tree  carbon  yield  (TCY).  The 
error  distributions  arising  from  all  lower  levels  of  derivation  will  have  to  be 
analyzed  outside  of  this  generalized  framework. 

In  the  example  of  the  TCY  (in  the  US  carbon  budget),  its  quantity  is 
not  measured  but  derived  from  timber  inventory  information  that  is  based  on 
measurements  of  diameters,  heights,  ages,  densities,  species  compositions,  etc., 
with  associated  sampling  errors,  measurement  errors  and  other  errors.  Those 
errors  are  propagated  by  applied  formulas  and  assumptions  that  are  used  to 
derive  TCY,  and  the  result  associates  certain  error  distributions  that  have 
to  be  derived  or  estimated  outside  of  the  generic  carbon  budget  uncertainty 
analysis  framework.  Exact  derivation  of  the  error  distributions  for  TCYs  would 
depend  on  designs  and  assumptions  of  each  individual  inventory  and  models 
applied  to  the  measured  quantities.  For  the  purpose  of  this  analysis  these  error 
distributions  have  to  be  based  on  expert  opinion. 

The  carbon  budget  uncertainty  analyses  are  considered  here  as  two  poten- 
tial studies.  The  first  study  is  the  uncertainty  analysis  of  the  carbon  pools 
at  different  aggregation  levels,  i.e.,  analysis  of  uncertainty  associated  with  a 
carbon  inventory  at  one  point  in  time  that  summarizes  a  specified  number  of 


651 


carbon  pools.  The  second  study  is  the  uncertainty  analysis  of  the  carbon  flux 
estimation,  i.e,  estimation  of  uncertainties  associated  with  estimated  changes 
in  the  carbon  pools  over  time. 

A  third  potential  area  of  effort  could  concentrate  on  analysis  of  uncertainty 
associated  with  estimation  of  a  periodic  carbon  flux  based  on  two  given  carbon 
inventories  conducted  at  different  times.  This  approach  would  involve  analysis 
of  conditional  uncertainties  depending  on  a  knowledge  of  statistical  distribu- 
tions of  uncertainties  associated  with  each  of  the  two  available  inventories. 
These  would  likely  be  correlated  and  potentially  would  not  have  normally  dis- 
tributed errors.  Nevertheless,  this  problem  could  also  be  addressed  through 
applications  of  Monte  Carlo  simulations,  though  the  results  would  not  be  easily 
generalized  due  to  the  conditional  aspect  of  such  approach. 

Definition  of  the  Carbon  Pools  and  their  Budget 

A  basic  entity  of  any  carbon  budget  is  a  carbon  store  at  a  given  level  of 
aggregation.  This  store  is  defined  by  its  level  of  aggregation,  denomination, 
area,  and  yield,  i.e.,  quantity  per  unit  area.  An  amount  of  carbon  in  a  store 
at  a  given  aggregation  level  and  for  a  given  denomination  is  calculated  for 
this  store  as  a  product  of  the  area  and  yield  of  this  store.  If  at  the  lowest 
level  of  aggregation  in  detailed  models  a  store  consists  of  sub-units  of  carbon 
storage  within  one  basic  unit  area,  i.e.,  basic  carbon  pools  having  defined 
individual  yields,  the  store  yield  is  then  simply  a  product  of  the  store  area 
and  its  respective  sum  of  pertaining  pool  yields.  All  of  the  carbon  stores  are 
parts  of  corresponding  higher  aggregation  stores  and  they  are  obtained  from 
adding  up  sums  of  lower  aggregation  stores  of  the  same  category.  Carbon 
stores  can  be  represented  within  the  analysis  framework  on  an  arbitrary  level 
of  aggregation.  An  example  of  a  carbon  store  at  high  level  of  aggregation  is 
tropical  forests  in  Amazonia.  An  example  of  a  carbon  store  on  a  low  level  of 
aggregation  is  live  tree  carbon  (TCY)  in  a  monospecific  stand  under  specific 
management  practices  in  one  management  unit  and  at  one  age. 

A  carbon  store  on  a  specific  aggregation  level  /  is  always  defined  as  simply 
a  sum  of  carbon  stores  within  the  same  category,  m,  at  /  —  1  level  of  aggrega- 
tion, i.e.,  C(/)m  =  5Zi  ('~1)m  Y(i-i)  where:  C(/)m  is  carbon  storage  at  the 
aggregation  level  (/)  in  the  category  m;  iV(/_i)  is  number  of  carbon  stores,  or 
pools,  at  level  (/  —  1)  in  the  category  m;  is  area  of  carbon  store  m  at 
/—I  aggregation  level;  and  Y(i-i)  is  the  yield  of  carbon  store  m  at  aggregation 
level  /-  1. 

The  rule  described  by  this  equations  is  applied  recursively  to  the  carbon 
stores.  If  the  areas  or  yields  on  a  level  (/  —  1)  are  not  defined  they  are  computed 
from  level  (/  —  2).  If  the  carbon  quantities  at  level  (/  —  2)  are  not  defined 
they  are  computed  from  level  (/  —  3),  and  so  on,  until  a  level  k  at  which 
the  quantities  of  carbon  are  defined  as  the  basic  model  input.  At  that  point 
the  summations  can  be  brought  up  gradually  from  level  k  to  level  (/).  Thus 
any  carbon  budget  is  defined  as  a  sum  of  products  between  areas  and  their 
yields  from  the  lowest  aggregation  level  of  carbon  pools,  even  though,  for 
stores  composed  of  sums  of  basic  carbon  pools  within  the  store  area  C(/)m  = 

X^i~1)m  X!j=i1)m  ^(J-i)OTi^</-i)mi  .?  where:  n  is  the  number  of  basic  pools  in  the 


652 


lowest  aggregation  level  store;  and  K(/_i)  is  a  yield  for  pool  j  in  category  m 
at  level  (/ -  1). 

Carbon  Budget  Marginal  Uncertainties 

First  order  marginal  uncertainties  around  the  carbon  budget  are  defined 
here  by  the  partial  derivatives  of  the  sum  of  basic  carbon  pools  within  the  store 
area  with  respect  to  its  input  variables  A  and  V ,  i.e.,  ^(0™   _  ^"l_'~1)m  t 

and  ~vdC{l)m —  =  An-i)  ... 

It  is  clear  from  the  first  order  marginal  uncertainties  that  the  second  order 
of  marginal  sensitivities  in  accordance  with  the  general  carbon  budget  scheme 
defined  here  would  be  equal  to  zero  for  both  area  and  stocking  related  errors. 
The  non-dimensional  quantities  in  the  first  order  linear  marginal  uncertainties 
are  defined  as  relative  values: 


(0m,*/  Wm*.     LJ=i     Lj=i  A(i-i)miy{i-i)mij 


and 


dY(,  u       I  Yd  \\  \-Nv-i)m  vn('-1}^  a  v 


(1) 


(2) 


The  total  error  of  the  carbon  budget  of  category  m  at  level  /  is  defined  as 
Y(i-i)mit  +-^(/-i)m     ^  we  define  the  total  error  in  the  carbon  budget 
at  /  level  of  aggregation  and  m  category  as  AC(/)m  and  the  errors  associated 
with  inputs  of  the  components  as  AA(/_i)m.  and  AY(i-i)m.  ,  then: 

N(l-Vm    (  n('-l)m  n('-l)m  \ 

AC„)m=   E    K-iU   E  .  +  AA()_1)mj  £  ^-DmJ  (3) 

1=1     \  i=i  i=i  / 

In  relative  terms  the  total  error  of  the  carbon  budget  of  category  m  at  level 
/  is  defined  as: 


Ac„)m  _      (^-u,  E^T1'-  AK(,-1)m„ + a^_,u  £;^>- 

c«>~  ~  Erii-,,"E^1,"A(,_1)>i>r(1_1Uj 

(4) 

Using  the  marginal  uncertainties  for  uncertainty  analysis  based  on  formal 
sensitivity  analysis  is  the  most  effective  way  to  treat  the  problem  of  calculating 
the  uncertainties  in  the  responses  of  carbon  budgets  simulations  due  to  uncer- 
tainties of  the  input  parameters.  This  method  is  effective  and  relatively  simple 
for  implementation  and  it  can  be  used  with  changing  input  error  assumptions. 
It  is  also  suitable  for  identifying  most  important  sources  of  errors. 


653 


ERROR  ANALYSIS  OF  CARBON  FLUX  ESTIMATION 


It  is  assumed  here  that  the  carbon  flux  is  computed  from  simulations  of 
changes  in  carbon  budget  over  time  rather  than  from  differences  between  car- 
bon budgets  at  different  times.  In  such  a  dynamic  approach  a  simulation  of 
error  propagation  in  carbon  flux  computations  should  be  responsive  to  initial 
input  values  and  errors  associated  with  them  as  well  as  to  potential  devia- 
tions from  any  underlined  growth  patterns.  Since  none  of  the  responses  can 
be  readily  simulated  using  yield  tables  the  simulation  model  should  be  based 
only  on  continuous  functions. 

To  facilitate  the  error  propagation  of  the  simulated  system,  functions  sim- 
ulating changes  in  carbon  pools  over  time  should  be  self-referencing.  That 
is,  they  should  be  based  on  dynamic  equations  or  difference  equations  that 
are  driven  by  initial  conditions  of  the  independent  and  dependent  variable  to- 
gether with  the  independent  variable  so  that  estimated  inventory  errors  can 
be  directly  incorporated  into  a  simulation  inputs. 

Development  of  the  dynamic  equations  is  necessary  for  those  carbon  bud- 
gets that  are  currently  based  on  yield  tables.  An  example  here  is  the  US  forest 
C  budget  (Turner  et  al.  1995)  which  is  based  on  a  complete  forest  inventory, 
a  set  of  stand  level  C  budgets  which  specify  the  age-specific  pools  and  flux 
associated  with  C  in  5  pools  (tree,  understorey,  forest  floor,  soil  and  woody 
debris),  and  a  specified  forest  harvest. 

In  this  example,  the  model  is  based  on  static  carbon  yield  tables.  Therefore 
the  first  step  in  computation  of  the  carbon  flux  error  propagation  is  to  develop 
a  set  of  dynamic  equations  replacing  the  static  yield  tables.  All  the  steps  for 
adopting  this  framework  to  the  Monte  Carlo  error  analysis  approach  can  be 
outlined  as  follows: 

1.  For  each  stand  level  C  budget,  replace  the  age-specific  pools/tables  and 

flux  values  by  equations  (e.g.,  tree  C  vs  age). 

2.  Identify  the  important  sources  of  error  within  the  algorithms  that  are  used 

to  create  the  stand  level  C  budgets. 

3.  Determine  an  appropriate  formulation  of  the  error  distribution  for  each 

case  (either  from  the  literature  or  using  an  assumption). 

4.  Create  the  mechanism  within  the  simulation  package  to  vary  the  matrix  of 

error  estimates  for  successive  Monte  Carlo  runs. 

5.  Run  the  pools  and  NEP  flux  calculation  multiple  times  using  the  varying 

values  for  specific  pools,  ratios,  area,  fluxes,  etc. 

6.  In  the  case  of  the  transition  flux,  determine  an  appropriate  error  distribu- 

tion for  the  areas  involved  and  perform  the  Monte  Carlo  simulation. 

7.  Examine  Monte  Carlo  output  for  specific  carbon  pools  and  fluxes,  for  dif- 

ferent levels  of  aggregation,  and  for  the  whole  system. 

Only  the  first  step  from  the  above  list  has  been  addressed  here.  Devel- 
opment of  equations  to  replace  tables  is  necessary  to  be  able  to  consistently 


654 


propagate  errors,  examine  the  influence  of  changes  in  simulated  trends  on  the 
error  propagation,  and  streamlining  the  computational  procedures  that  would 
be  involved  in  numerous  repetitions  of  all  the  computations  involved  in  Monte 
Carlo  simulations.  Furthermore,  to  examine  the  potential  impact  of  inventory 
errors  on  the  carbon  budget  simulations  it  is  advantages  to  form  the  yield  equa- 
tions as  self-referencing  functions  or  dynamic  equations.  Such  equations  have 
been  developed  for  all  the  US  carbon  budget  yield  tables  and  are  presented  in 
the  following  sections  of  this  paper. 

Example  of  Flux  Estimation:  US  Carbon  Budget 

The  Tables 

The  carbon  yield  over  age  data  used  in  the  US  carbon  budget  consisted  of 
tables  with  carbon  yields  of  five  basic  categories  of  carbon  pools:  i)  live  tree 
carbon;  ii)  soil  carbon;  iii)  forest  floor  carbon;  iv)  understory  carbon;  and 
v)  coarse  woody  debris  carbon.  Included  were  tables  for  both  yields  and  yearly 
increments  for  all  these  categories.  The  tables  contained  data  for  total  carbon 
yields  and  growth  increments  for  total  carbon,  i.e.,  sums  of  the  five  categories, 
and  for  merchantable  wood  volume. 

Based  on  species  associations,  the  tables  were  aggregated  into  50  groups 
and  plotted  for  visual  inspections  and  preliminary  analysis  of  potential  growth 
patterns.  Some  tables  showed  drastic  inconsistencies  in  growth  patterns  sug- 
gesting that  either  the  grouping  criteria  were  too  general  or  the  tables  are 
deficient.  This  means  that  the  future  work  should  address  the  problem  of 
growth  pattern  inconsistencies  within  species  grouping  in  the  US  carbon  yield 
tables  and  more  detailed  work  will  have  to  be  done  to  define  more  consistent 
clustering  of  the  carbon  yield  patterns. 

Ranges  in  stand  age  vary  from  0-100  to  0-180  years  for  different  species, 
management  units,  and  management  practices.  Some  tables  include  before 
and  after  thinning  values.  There  are  about  500  sets  of  tables  for  about  50 
combinations  of  species.  After  separating  the  tables  with  before  and  after 
thinning  values  there  are  over  6000  tables  describing  the  carbon  changes  as  a 
function  of  age,  in  a  >  1  Megabyte  file  with  a  total  of  about  200,000  discrete 
numbers,  that  would  have  to  be  interpolated  for  any  ages  not  included  in  the 
tables. 

Given  that  every  table  for  each  single  carbon  pool  contains  about  40  num- 
bers that  may  need  further  interpolations,  developing  a  set  of  equations  with  a 
unique  set  of  coefficients  for  each  table  would  be  a  great  advance  in  streamlin- 
ing the  calculations  involved  in  the  US  carbon  budget.  The  equations  would 
not  only  reduce  the  number  of  involved  references,  but  also  would  allow  for 
more  accurate  and  continuous  computations  of  carbon  quantities  over  time. 
Furthermore,  dynamic  equations,  unlike  static  tables,  or  for  that  matter  even 
static  equations,  allow  for  implementation  of  error  propagation  analysis  in  flux 
calculations  and  for  clustering  of  the  tables  into  common  growth  patterns  even 
at  variable  productivity  levels  between  different  tables. 


655 


The  Dynamic  Equations 


Five  types  of  dynamic  equations  have  been  developed  for  the  US  Carbon  Bud- 
get Flux  estimation  for  the  five  different  pools.  Nonlinear  regression  fittings  of 
these  equations  varied  in  performance  and  the  number  of  needed  sets  of  coef- 
ficients to  describe  all  the  relevant  tables.  The  best  performance  was  achieved 
with  the  equation  for  forest  floor  carbon;  only  one  set  of  coefficients  was  re- 
quired to  fit  well  all  of  the  forest  floor  tables. 

The  developed  dynamic  equations  are  truly  base-age  invariant  and  they 
are  based  on  different  base  equations  and  different  assumptions  of  the  cross- 
sectional  changes  between  different  tables.  The  analytical  derivation  details  of 
all  equations  are  beyond  the  scope  of  this  paper  and  are  described  in  another 
manuscript.  Graphs,  coefficients  and  statistical  summaries  of  the  relevant  re- 
gression analysis  for  the  calibration  of  the  dynamic  equations  are  available  on 
request  from  the  principal  author  of  this  paper.  The  forms  of  the  developed 
dynamic  equations  are  described  below  under  subsequent  headings  denoting 
the  five  relevant  carbon  pools  of  the  US  Carbon  Budget.  The  following  sym- 
bols are  used  in  the  illustrated  below  dynamic  equations:  i)  a,  /?,  . . .  are  the 
equation  estimable  coefficients;  ii)  t  is  the  age  of  prediction;  iii)  tQ  the  age  of 
reference  at  which  a  value  of  carbon  is  known  and  used  as  the  equation  input; 
iv)  ya  is  the  known  value  of  carbon  at  age  t0  that  is  used  as  the  equation  initial 
value;  and  v)  Y  is  the  carbon  yield  at  age  given  the  initial  value  of  carbon 
at  age  tQ  equal  to  yQ. 

Forest  Floor  Carbon  —  The  growth  and  yield  of  carbon  in  forest  floor  data 
was  the  most  consistent  for  all  tables  and  the  developed  model  was  able  to 
describe  the  carbon  growth  and  yield  for  all  tables  (R2  >  99.93%)  with  a  single 
set  of  coefficients.  The  curves  generated  by  this  equation  have  different  values 
at  origin  and  they  are  polymorphic  with  variable  asymptotes.  The  equation 
has  the  following  form: 

Y(t)  -    -  (/?  +  *g)  (*7  +  Vo)     I  Vo  (1  +  ^  =  7  (S\ 

u        (<$/? +tg  +  <^r  i+/3/*8  +  i/tf  {) 

Live  Tree  Carbon  —  Fifty  sets  of  coefficients  were  fitted  to  all  the  predefined 
groups  of  tree  live  carbon.  The  curves  generated  by  this  equation  are  equal 
to  zero  at  origin  and  they  are  polymorphic  with  variable  asymptotes.  The 
consistency  of  growth  patterns  was  fairly  good  with  few  minor  exceptions  and 
the  fitted  model  explained  much  of  the  variation  between  different  yield  tables 
and  it  has  the  following  form: 


t 


a 


(2  +  *  +  ^ +  (-£  +  »)') 


Coarse  Woody  Debris  Carbon  —  Like  the  live  tree  carbon  the  coarse 
woody  debris  carbon  growth  and  yield  is  described  by  one  equation  with  50 


656 


sets  of  coefficients.  The  fits  are  not  as  good  as  for  the  live  tree  carbon  but 
were  good  enough  to  justify  the  grouping  of  the  tables  into  common  patterns 
described  by  the  following  equation: 

Y  =  at  -  8t3  +  (y0  -  at0  +  A>)  l^B±J^  (7) 

Understory  Carbon  -  The  carbon  dynamics  of  the  understory  were  very 
inconsistent.  The  changes  over  time  were  not  continuous  and  seemed  to  consist 
of  arbitrary  numbers.  No  generalizations  could  be  developed  within  the  time 
frame  available  for  this  analysis.  The  understory  carbon  dynamics  equation 
had  to  be  fitted  to  each  table  individually  resulting  in  647  sets  of  coefficients. 
Even  with  individually  fitted  coefficients  the  regressions  did  not  explain  much 
of  the  variation  seemingly  present  in  this  variable.  All  647  sets  of  the  coeffi- 


cients relate  to  one  equation  of  relatively  low  complexity  Y(t)  = 


Soil  Carbon  —  The  yields  of  soil  carbon  were  constants  in  almost  all  tables 
and  accordingly  their  increments  were  zero.  In  a  few  tables,  the  soil  carbon 
was  increasing  with  a  constant  growth  rate.  A  simple  difference  equation  with 
only  one  coefficient  (Y(t)  =  y0  jjjjgr )  can  describe  this  behaviour  without  a 
need  for  regression  analysis. 

SHORT  DISCUSSION  AND  CONCLUSION 

The  described  study  aimed  towards  developing  a  generic  framework  for 
error  analysis  on  national  level  carbon  budgets  and  flux  tracking  models.  A 
number  of  steps  have  been  outlined  for  this  development  and  two  important 
steps  have  been  advanced.  The  outlined  points  of  consideration  for  potential 
Monte  Carlo  applications  may  be  helpful  in  planning  and  undertaking  similar 
studies. 

The  synopsis  of  the  formal  sensitivity  analysis  through  analytical  definition 
of  marginal  uncertainties  potentially  can  improve  efforts  of  uncertainty  analysis 
through  sensitivity  analysis  in  comparison  to  these  attempted  in  the  past, 
e.g.,  Kurz  et  al.  (1994).  The  conceptualized  generalization  of  a  carbon  budget 
modeling  framework  could  be  used  to  develop  a  generic  carbon  budget  dynamic 
simulation  software  that  would  be  used  for  both  formalized  sensitivity  analysis 
and  Monte  Carlo  simulations  of  different  hypothetical  carbon  budgets. 

Finally,  the  development  of  the  dynamic  equations  for  the  US  carbon  bud- 
get illustrates  potential  gains  in  the  development  of  such  equations  with  regard 
to  reducing  the  amount  of  information  which  must  be  stored  and  increasing 
the  flexibility  in  model  testing  and  applications.  Development  of  these  equa- 
tions makes  possible  error  propagation  analysis  in  carbon  flux  calculations  and 
model  behaviour  investigation  that  could  not  be  attempted  with  tables. 

LITERATURE  CITED 

Cunia,  T.  1985.  On  volume  tables  and  their  contribution  to  the  error  of  forest 
inventory  volume  estimates.  P.  205-222  in  Proc.  National  Forest  Inven- 
tory in  Europe  Workshop.  University  of  Freiburg-Faculty  of  Forestry, 
Freiburg,  Federal  Republic  of  Germany. 


657 


Gertner,  G.,  and  M.  Kohl.  1992.  An  assessment  of  some  nonsampling  errors 
in  a  national  survey  using  an  error  budget.  Forest  Science  38:  525-538. 

Intergovernmental  Panel  on  Climate  Change.  1995.  IPCC  Guidelines  for 
National  Greenhouse  Gas  Inventories.  3  vols.,  IPCC,  Bracknell,  UK. 

Kauppi,  P.  E.,  K.  Mielikoinen,  and  K.  Kuusela.  1992.  Biomass  and  carbon 
budget  of  European  forests,  1971  to  1990.  Science  256:  70-74. 

Kolchugina,  T.  P.,  and  T.  S.  Vinson.  1993.  Carbon  sources  and  sinks  in  forest 
biomes  of  the  Soviet  Union.  Global  Biogeochemical  Cycles  7:  291-304. 

Kurz,  W.  A.,  M.  J.  Apps,  T.  M.  Webb,  and  P.  J.  McNamee.  1992.  Informa- 
tive report.  The  carbon  budget  of  the  Canadian  forest  sector:  Phase  I, 
NOR-X-326.  Forestry  Canada,  Edmonton,  Alberta,  Canada. 

Kurz,  W.  A.,  M.  J.  Apps,  S.  J.  Beukema,  and  T.  Lekstrum.  1994.  20th 
century  carbon  budget  of  Canadian  forests.  Tellus  47  B:  170-177. 

Maclaren,  J.  P.,  and  S.  J.  Wakelin.  1991.  Forestry  and  forest  products  as 
a  carbon  sink  in  New  Zealand.  FRI  Bulletin.  Number  162.  Ian  Bryce, 
Rotorua,  New  Zealand. 

Makundi,  W.,  J.  Sathaye,  and  0.  Masera.  1992.  Carbon  emission  and  seques- 
tration in  forest:  Case  studies  from  seven  developing  countries,  Vol.  1: 
Summary.  Report  No.  LBL-32199,  Lawreace  Berkeley  Laboratory,  Uni- 
versity of  California,  Berkeley,  CA. 

Robinson,  J.M.  1989.  On  uncertainty  in  the  computation  of  global  emissions 
from  biomass  burning.  Climatic  Change  14:243-262. 

Turner,  D.  P.,  J.  J.  Lee,  G.  J.  Koerper,  and  J.  R.  Barker,  editors.  1993. 
The  forest  sector  carbon  budget  of  the  United  States:  carbon  pools  and 
flux  under  alternative  policy  options,  EPA/600/3-93/093.  United  States 
Environmental  Protection  Agency,  Environmental  Research  Laboratory, 
Corvallis,  Oregon,  USA. 

Turner,  D.  P.,  G.  J.  Koerper,  M.  E.  Harmon,  and  J.  J.  Lee.  1995.  A  car- 
bon budget  for  forests  of  the  conterminous  United  States.  Ecological 
Applications  5(2):  421-436. 

BIOGRAPHICAL  SKETCH 

Chris  J.  Cieszewski  is  a  Forest  Biometrician  with  Man  Tech  Environmental 
Research  Services  Corp.,  Corvallis,  OR.  He  holds  a  Ph.D.  in  Forest  Biomet- 
rics from  University  of  Alberta,  has  been  working  in  his  field  since  1986  and 
specialises  in  development  of  mixed-effects  base- age  invariant  models. 

David  P.  Turner  is  a  Forest  Ecologist  with  ManTech  Environmental  Re- 
search Services  Corp.,  Corvallis,  OR.  He  holds  a  Ph.D.  in  Botany  from  Wash- 
ington State  University  and  specialises  in  the  area  of  forest  nutrient  cycling. 

Don  L.  Phillips  is  a  Research  Biologist  with  the  U.S.  EPA  National  Health 
Sz  Environmental  Effects  Laboratory,  Corvallis,  OR.  He  holds  a  Ph.D.  in  Bi- 
ology from  Utah  State  University,  and  has  worked  in  EPA's  climate  change 
research  program  since  1988. 


658 


Monte  Carlo  Simulations  of  Nonlinear 
Size-Age  Relationships 

Ronald  E.  McRoberts1 


Abstract. -Monte  Carlo  simulations  are  an  accepted  and  useful  procedure 
for  determining  the  effects  of  uncertainty  in  input  variables  and 
parameters  on  the  uncertainty  of  derived  output  variables.  Monte  Carlo 
procedures  are  particularly  useful  for  growth  model  applications  when  the 
model  is  nonlinear  in  the  parameters.  For  example,  a  nonlinear  growth 
model  may  be  fit  to  variables  derived  from  size-age  data,  and  then  Monte 
Carlo  simulations  may  be  used  to  estimate  the  parameter  covariance 
matrix  and  a  confidence  interval  resulting  from  repeatedly  applying  the 
growth  model  beginning  with  a  set  of  initial  conditions.  However,  when 
simulated  values  of  size  variables  contain  randomly  generated 
components,  negative  growth  may  occur.  An  example  using  height-age 
data  and  the  Weibull  growth  model  is  used  to  illustrate  and  compare 
approaches  to  this  problem  and  to  justify  an  assertion  that  it  may  be 
difficult  to  precisely  predict  future  height  


INTRODUCTION 

Monte  Carlo  simulation  techniques  have  been  recognized  as  a  valid  method  for 
determining  the  effects  of  uncertainty  in  input  variables  and  parameters  on  the 
uncertainty  of  derived  output  variables  (McRoberts  1996,  1995,  1994,  Mowrer 
1994,  1992,  Makela  1988,  Gertner  and  Dzialowy  1984).  They  are  particularly 
useful  for  estimating  parameter  covariance  structures  and  for  propagating 
uncertainty  through  nonlinear  growth  models  when  linear  approximations  perform 
poorly,  when  the  mathematical  formulation  is  complex,  and  when  predictor 
variables  depend  on  predicted  variables  at  a  previous  stage. 

The  underlying  objective  is  to  estimate  the  bias  and  uncertainty  in  size 
predictions  obtained  from  repeatedly  applying  an  annual  growth  model  beginning 
with  an  observed  initial  condition.  Calibration  data  for  the  growth  model  consist 
of  a  series  of  size-age  observations  for  a  typical  tree  whose  future  size  is  to  be 
predicted.  Nonlinear  regression  is  used  to  estimate  the  model  parameters,  and 
Monte  Carlo  simulations  are  used  to  estimate  the  covariance  structure  of  the 
parameter  estimates  and  confidence  intervals  for  future  size  predictions.  Two 
difficulties  arise  when  applying  simulations  to  this  problem:  (1)  accommodating 
the  repeated  measures  nature  of  the  data  when  simulating  observations  and  when 
calibrating  the  growth  model;  and  (2)  dealing  with  negative  growth  that  occurs 


Mathematical  Statistician,  North  Central  Forest  Experiment  Station,  1992  Folwell  Ave.,  St.  Paul,  MN  55108. 


659 


when  a  simulated  size  observation  is  less  than  a  simulated  size  observation  at  a 
previous  age  due  to  randomly  generated  residuals. 


DATA 


The  calibration  data  consist  of  either  13  or  14  height- age  observations  for  each 
of  six  co-dominant  yellow  birch  trees  from  the  same  forest  stand.  The  data  for 
this  study  are  a  subset  of  the  yellow  birch  stem  analysis  data  collected  by 
Carmean  (1978)  and  used  to  construct  northern  hardwoods  site  index  curves. 
They  were  collected  from  a  plot  in  a  fully  stocked,  even-aged  stand  on  the 
Chequamegon  National  Forest  in  northern  Wisconsin.  Annual  rings  were  counted 
on  disks  cut  at  each  section  point.  For  Carmean's  analyses,  a  height-age  graph 
was  constructed  for  each  tree  and  was  examined  for  signs  of  early  suppression,  top 
breakage,  or  dieback.  Such  disturbed  trees  were  discarded  as  were  those  from 
even-aged  stands  whose  age  range  for  dominant  and  codominant  trees  exceeded 
10  years. 

These  height-age  data  are  characterized  as  longitudinal,  repeated  measures  data: 
(1)  they  were  obtained  by  repeatedly  observing  the  ages  of  the  same  trees  at 
different  heights;  (2)  they  were  necessarily  ordered  in  time  and  position  and 
allowed  no  randomization;  and  (3)  the  trees  from  which  they  were  obtained  were 
assumed  to  constitute  a  representative  sample  from  the  population  of  dominant  and 
co-dominant  trees.  The  serial  correlation  exhibited  by  longitudinal,  repeated 
measures  data  due  to  the  lack  of  randomization  must  be  accommodated  in 
parameter  estimation  routines  if  unbiased  confidence  intervals  are  to  be  obtained. 


METHODS 


Height-age  model 

McRoberts  (1996,  1995)  has  shown  that  the  height-age  relationship  for  these 
data  is  adequately  described  using  a  Weibull  model, 

a  BX& 

Y-fta-eP^  )+£i  [1] 

where  Y  is  height,  x  is  age,  and  ft =Oi,p2,P3)  is  a  parameter  vector  to  be  estimated. 
The  residuals,  €,  are  distributed  N^^A)  with  elements,  of  the  correlation 
matrix,  A,  structured  as 

A^p1*^1'  [2] 


where  p  is  the  annual  serial  correlation  and  a2  is  the  residual  variance. 

To  obtain  the  height-age  relationship  for  a  typical  tree,  population  parameters 
for  Eq.  [1]  were  estimated  using  a  nonlinear  mixed  effects  modeling  approach. 
Models  representing  the  height-age  relationship  for  a  population  of  trees  contain 


660 


both  fixed  effects  (population  parameters)  and  random  effects  (individual 
parameters)  and  are  described  as  mixed  effects  models.  To  simultaneously 
accommodate  the  mixed  effects  and  longitudinal,  repeated  measures  nature  of  the 
data,  the  first-order  linearization  procedure  for  nonlinear  mixed  effects  models 
described  by  Davidian  and  Giltinan  (1995)  was  used.  Two  assumptions  underlie 
this  procedure:  (1)  the  same  model  form,  but  with  possibly  different  parameter 
values,  is  adequate  to  describe  the  height-age  relationship  for  all  trees;  and  (2)  the 
same  residual  variance,  correlation  structure,  and  correlation  parameter  value  apply 
for  all  trees.  The  procedure  applies  least  squares  procedures  to  estimate  the  fixed 
and  random  model  parameters  and  maximum  likelihood  procedures  to  estimate  the 
common  residual  variance,  a2,  and  annual  serial  correlation,  p. 

The  residuals  for  estimated  individual  tree  height-age  curves  exhibited 
homogeneity  of  variance.  This  result  is  consistent  with  the  assumption  that  the 
standard  deviation  of  growth  is  proportional  to  the  expectation  of  growth,  because 
height  increments  between  adjacent  observations  were  always  approximately  four 
feet.  An  estimate,  cCT,  of  the  proportion  was  obtained  by  a  second  application  of 
the  nonlinear  mixed  effects  model  routine  using  a  covariance  structure  based  on 
this  assumption  and  an  assumption  of  annual  height  growth  correlation,  p. 

Height  growth  model 

Three  variables  were  derived  from  the  original  height-age  observations:  (1) 
average  growth,  the  difference  in  height  between  adjacent  sections  divided  by  the 
corresponding  difference  in  age;  (2)  average  height  of  adjacent  sections;  and  (3) 
average  age  of  adjacent  sections.  These  variables  were  used  to  develop  a  model 
for  predicting  average  annual  growth  as  a  function  of  average  age  and  average 
height.  Because  the  Weibull  model  has  been  shown  to  adequately  describe  the 
height-age  relationship  for  these  data,  height  growth  was  modeled  using  a 
differential  form  of  Eq.  [1], 

Yi=a1xlia2(x2i-a3)+ei  [3] 


where  Y  is  average  height  growth,  xt  is  average  age,  x2  is  average  height,  and 
a=(a1,oc2,ot3)  is  a  parameter  vector  to  be  estimated.  The  parameters,  a,  of  Eq.  [3] 
correspond  to  the  parameters,  P,  of  Eq.  [1]  as  follows:  a1=P2(33,  02=P3-1,  a3=Pi. 
The  residuals,  €,  are  distributed  N(0,V)  where  the  elements  of  V  are  calculated  on 
the  basis  of  the  following  assumptions:  (1)  variance  is  proportional  to  the  square 
of  expected  growth;  (2)  variance  is  inversely  proportional  to  the  number  of  years 
over  which  average  growth  is  calculated;  and  (3)  the  correlation  among 
observations  is  derived  from  Eq.  [2]. 

Confidence  Intervals 

A  6-step  Monte  Carlo  procedure  was  used  to  estimate  an  approximate  95% 
confidence  interval  around  the  estimated  height-age  curve  obtained  from  repeated, 


661 


annual  applications  of  the  growth  model.  To  estimate  these  confidence  intervals, 
random  numbers  were  generated  to  simulate  uncertainty  from  two  sources:  (1) 
variability  in  the  estimates  of  model  parameters,  and  (2)  residual  variability  around 
the  curves  resulting  from  the  parameters  estimates. 

Step  1.  Selection  of  a  typical  tree 

Estimates  of  the  population  parameters,  /3,  and  the  common,  individual  tree 
estimates  of  a2,  cc,  and  p  were  used  to  generate  data  for  a  typical  tree. 

Step  2.  Simulation  of  height-age  observations 

Four-foot  increments,  beginning  at  8  feet  and  ending  at  64  feet,  were  selected, 
and  Eq.  [1]  was  solved  using  the  population  parameter  estimates  selected  in  Step 
1  to  determine  the  corresponding  expected  ages.  A  multivariate  normal  vector  of 
residuals,  €,  was  randomly  generated  using  the  common  values  of  a2  and  p  and 
was  added  to  the  selected  heights  to  simulate  a  height-age  series.  Random 
selection  of  residuals  may  cause  simulated  height  at  a  subsequent  age  to  be  less 
than  simulated  height  at  a  previous  age.  Three  approaches  to  this  negative  height 
growth  problem  were  implemented:  (Approach  A)  ignore  the  problem;  (Approach 
B)  reject  simulated  height-age  series  exhibiting  negative  height  growth;  and 
(Approach  C)  reject  simulated  height-age  series  whenever  the  absolute  value  of 
a  residual  is  greater  than  expected  height  growth.  Average  height  growth,  average 
age,  and  average  height  were  calculated  from  these  simulated  data. 

Step  3.  Estimation  of  growth  model  parameters 

The  parameters  for  Eq.  [3]  were  estimated  from  the  simulated  average  annual 
data  calculated  in  Step  2  with  weighted  nonlinear  least  squares  (WNLS).  The 
values  of  c0  and  p  selected  in  Step  1  were  used  to  calculate  a  weight  matrix  that 
was  constant  for  all  simulations. 

Step  4.  Simulation  of  annual  height  predictions 

Separate  50-year  series  of  annual  height-age  observations  were  simulated 
beginning  with  the  first  age  calculated  in  Step  2,  and  beginning  with  ages  in  ten 
year  increments  thereafter.  For  each  year  in  each  50-year  series,  height  was 
calculated  as  the  sum  of  previous  height,  predicted  annual  height  growth,  and  a 
randomly  generated  residual.  Predicted  height  was  obtained  from  Eq.  [3]  using 
the  parameters  estimated  in  Step  3.  The  residuals  were  multivariate  normal  with 
serial  correlation,  p,  and  standard  deviations  calculated  as  the  product  of  ca  and 
predicted  height  growth.  The  simulated  height-age  series  were  saved  for 
calculation  of  confidence  intervals  in  Step  6.  Additional  values  of  cc  (ca=0.50, 
cCT=0.60)  were  also  used  to  determine  the  effect  of  this  parameter  on  bias  and 
uncertainty.  The  same  three  approaches  to  negative  height  growth  described  in 
Step  2  were  implemented  in  the  simulation  of  these  annual  height-age 
observations. 


662 


Step  5.  Replication 

Steps  2-4  were  replicated  1000  times  for  each  approach  to  the  negative  height 
growth  problem  and  each  value  of  cCT. 

Step  6.  Construction  of  confidence  intervals 

For  each  year  in  the  50-year  series,  the  1000  annual  height  predictions  were 
ordered  from  smallest  to  largest,  and  those  corresponding  to  the  2.5,  50.0  and  97.5 
percentiles  were  selected.  The  curve  resulting  from  the  50th  percentile  predictions 
across  all  50  ages  approximates  a  median  curve,  while  the  other  two  percentiles 
jointly  approximate  a  95%  confidence  interval.  Finally,  an  expected  height-age 
curve  with  no  random  variability  added  was  calculated  using  Eq.  [3]  with 
parameters  derived  directly  from  the  population  parameters  selected  in  Step  1. 

Strictly  for  comparison  purposes,  1000  simulations  were  also  conducted  using 
an  ordinary  nonlinear  least  squares  (ONLS)  approach.  These  simulations  were 
similar  to  those  previously  described  with  the  following  exceptions:  (1) 
independent,  normal  residuals  with  variance,  a2,  were  used  in  Steps  2  and  4;  (2) 
the  negative  height  growth  problem  was  completely  ignored  in  Steps  2  and  4;  and 
(3)  ONLS  was  used  for  parameter  estimation  in  Step  3. 

RESULTS  AND  CONCLUSIONS 

Parameter  estimates  for  the  height- age  population  curve  (Fig.  1)  were  obtained 
using  the  nonlinear  mixed  effects  model  routine:  £=(82.0107 ,-0.0123, 1.1409); 
a2=1.92;  co=0.39;  and  p=0.67.  The  cumulative  distribution  of  standardized 
residuals  indicated  the  assumption  of  normal  residuals  was  justified. 

For  each  simulation  of  15  height-age  observations  in  Step  2,  there  were 
approximately  0.15  rejections  for  Approach  B  and  approximately  0.50  rejections 
for  Approach  C.  In  Step  4,  rejections  occurred  approximately  0.20  times  for  each 
simulation  of  50  annual  observations  for  Approach  B  and  approximately  0.50 
times  for  Approach  C.  With  respect  to  Approach  A,  confidence  intervals  for  both 
Approaches  B  and  C  exhibited  bias.  For  Approach  B,  both  the  median  curve  and 
lower  confidence  limit  exhibited  positive  bias,  while  for  Approach  C  the  median 
curve  was  unbiased,  but  the  confidence  interval  was  too  narrow.  Although  the 
magnitude  of  the  bias  was  not  pronounced  for  ca=0.39,  it  increased  with  larger 
values  of  ca  (Fig.  2). 

The  median  curve  for  the  ONLS  method  was  unbiased,  but  the  95%  confidence 
interval  was  too  large.  This  result  confirms  the  necessity  of  accommodating  serial 
correlation  and  heterogeneity  of  variance  when  using  Monte  Carlo  simulations  and 
when  estimating  parameters  from  longitudinal,  repeated  measures  data. 

The  results  suggest  that  for  ca  >0.40,  an  unpleasant  dilemma  may  arise:  either 
(1)  accept  statistical  bias  as  the  price  to  pay  for  biological  realism,  or  (2)  tolerate 
biologically  impossible  conditions  in  order  to  avoid  statistical  bias.    From  a 


663 


strictly  statistical  perspective,  however,  negative  height  growth  poses  no  inherent 
conceptual  difficulties.  Even  though  the  expected  curve  is  monotonically 
increasing,  the  underlying  statistical  assumptions  do  not  require  subsequent 
observations  to  be  larger  than  previous  observations;  the  only  requirement  is  that 
the  residuals  satisfy  the  distributional  assumptions.  Thus,  procedures  for 
constructing  confidence  intervals  that  properly  reflect  residual  uncertainty  and 
uncertainty  in  parameter  estimates  must  ignore  negative  height  growth. 

70  r 


0 1       ■       1       ■       '       ■       ■       ■       ■  '       ■       ■       ■  ■ 

0  10  20  30  40  50  60  70 

Age  (yr) 

Figure  1.  Height-age  relationships. 


664 


Overall,  by  the  fiftieth  year  of  predicted  height,  confidence  intervals  increased 
to  a  width  of  about  15-20  feet,  depending  on  the  initial  condition  (Fig.  3). 
Coefficients  of  variation  were  estimated  by  dividing  Va  the  width  of  the  confidence 
interval  by  predicted  height.  The  estimates  were  generally  greater  when  the  initial 
age  and  the  length  of  the  prediction  interval  were  small  (Table  1).  The  general 
guideline  that  coefficients  of  variation  ought  not  exceed  about  0.10  (Gertner  1994) 
suggests  that  for  some  situations  it  may  be  extremely  difficult  to  precisely  predict 
future  height  using  an  annual  growth  model.  In  addition,  the  results  obtained  in 
this  study  must  be  considered  nearly  optimal,  because  the  calibration  data 
represented  only  the  tree  to  which  the  growth  model  was  applied.  For  an  actual 
application,  the  calibration  data  would  include  variation  among  a  large  sample  of 
trees  as  an  additional  source  of  uncertainty  and  would  produce  much  larger 
coefficients  of  variation. 


Q  1        i        i        i        i   i        .        i        .        i        .        i       .  .        i        .        i        .        i        .  i 

0         10        20        30        40        50        60        70        80        90  100 

Age  (yr) 

Figure  3.  95%  confidence  intervals. 


Table  1.  Coefficients  of  variation. 


Initial  age 

(yr) 

10 

20 

Length  of  prediction 

(yr) 

30 

40 

50 

6 

0.17 

0.13 

0.11 

0.08 

0.07 

16 

0.09 

0.08 

0.07 

0.07 

0.06* 

26 

0.05 

0.05 

0.06 

0.05* 

0.06' 

36 

0.04 

0.05 

0  05* 

0.07* 

0.08' 

46 

0.03 

0.04* 

0.06' 

0.08* 

0.09* 

*  Beyond  range  of  calibration  data. 


665 


REFERENCES 

Carmean,  W.H.  1978.  Site  index  curves  for  northern  hardwoods  in  Northern 
Wisconsin  and  Upper  Michigan.  USD  A  For.  Serv.  Res.  Pap.  NC-160. 

Gertner,  G.Z.  1994.  A  quality  assessment  of  a  Weibull  based  growth  projection 
system.  In:  Growth  and  Yield  from  Estimation  from  Successive  Forest 
Inventories.  Proc.  IUFRO  Conf.,  Copenhagen,  Denmark,  June  14-17, 1993. 

Gertner,  G.Z.,  and  Dzialowy,  P.J.  1984.  Effects  of  measurement  errors  on  an 
individual  tree-based  growth  projection  system.  Can.  J.  For.  Res.  14:311- 
316. 

Makela,  A.  1988.  Performance  analysis  of  a  process-based  stand  growth  model 
using  Monte  Carlo  techniques.  Scand.  J.  For.  Res.  3:315-331. 

McRoberts,  R.E.  1996.  Estimating  variation  in  field  crew  estimates  of  site  index. 
Can.  J.  For.  Res.  (in  press). 

McRoberts,  R.E.  1995.  Assumptions  underlying  estimation  methods  for  nonlinear 
mixed  effects  models.  In:  Am.  Stat.  Assn.  1995  Proc.  Biom.  Section, 
August  13-17,  1995,  Orlando.  FL.  (in  press). 

McRoberts,  R.E.  1994.  Variation  in  forest  inventory  field  measurements.  Can. 
J.  For.  Res.  24:1766-1770. 

Mowrer,  H.T.  1994.  Monte  Carlo  techniques  for  propagating  uncertainty  through 
simulation  models  and  raster- based  GIS.  In:  Proceedings  of  the 
International  Symposium  on  the  Spatial  Accuracy  of  Natural  Resource  Data 
Bases,  May  16-20, 1994,  Williamsburg,  VA,  R.G.Congalton,  ed.  Am.  Soc. 
Photogrammetry  and  Rem.  Sens.  ppl79-188. 

Mower,  H.T.  1992.  Open- architecture  Monte  Carlo  uncertainty  analysis  for 

forest  stand  dynamics  Models.  In:  Research  on  Growth  and  Yield  with 
Emphasis  on  Mixed  Stands,  IUFRO  Centennial  Meeting  Session  of  Section 
4.01,  Berlin/Eberswalde,  German,  August  31-September  4, 1992.  pp  75-84. 

BIOGRAPHICAL  SKETCH 

Ron  McRoberts  is  a  member  of  the  Forest  Inventory  and  Analysis  unit  of 
the  North  Central  Forest  Experiment  Station,  USDA  Forest  Service,  where  he 
develops  individual  tree  growth  models.  His  interests  include  nonlinear  regression 
modeling,  repeated  measures,  and  propagation  of  variance. 


666 


Cross-correlations  among 
Single  Tree  Growth  Models 

Hubert  Hasenauer1,  Robert  A.  Monserud2, 
Timothy  G.  Gregoire3 


Abstract.  -  Single  tree  growth  and  yield  models  basically  consist  of  a 
number  of  equations  to  update  tree  parameters  over  time.  Although  it 
seems  reasonable  to  assert  that  these  equations  are  interrelated  from  a 
biological  standpoint,  it  is  customary  to  consider  them  independently  and 
apply  linear  or  nonlinear  regression  techniques  separately  rather  than 
jointly.  Using  more  than  7,500  Norway  spruce  (Picea  abies)  trees,  we 
compare  an  individual  tree  basal  area  increment  model,  a  height 
increment  model,  and  a  crown  model  using  least  square  methods 
separately  and  jointly  by  applying  three  stage  least  square  (3SLS) 
regression  techniques.  Results  indicate  a  strong  cross-equation  correlation 
between  the  basal  area  and  the  height  increment  model  and  the  height 
increment  and  the  crown  model.  This  suggests  that  the  use  of  joint 
regression  techniques  would  be  superior. 


INTRODUCTION 

Individual  tree  forest  growth  and  yield  models  usually  employ  a  set  of 
equations  to  describe  stand  development  over  time.  A  typical  single-tree  stand 
simulator  (Monserud  1975,  Wykoff  et  al.  1982,  Burkhart  et  al.  1987,  Hasenauer 
1994)  consists  of  different  equations  for  predicting  periodic  diameter  or  basal  area 
increment,  height  increment,  and  the  probability  of  mortality  for  each  sample 
tree.  These  equations  are  usually  developed  separately. 

From  a  biological  standpoint,  it  seems  reasonable  to  assert  that  the  change 
in  a  tree's  basal  area,  height,  and  risk  of  mortality  are  not  uncorrelated 
phenomena  (Dixon  et  al.  1990).  Depending  on  their  interrelationships,  joint 
estimation  of  the  models'  parameters  may  be  necessary  in  order  to  provide 
estimates  that  are  consistent,  or  it  may  be  desirable  in  order  to  provide  more 
precise  estimates  than  can  be  obtained  otherwise. 

Seminal  work  by  Aitken  (1934-35),  Haavelmo  (1943),  Theil  (1953), 


1  Biometrician,  InstitutfiirWaldwachstumsforschung,  Universitdt  fur  Bodenkultur,  Wien,  Austria. 

2  Biometrician,  Intermountain  Research  Station,  USDA  Forest  Sendee,  Moscow,  ID  83843,  USA. 

3  Biometrician,  Dept.  Forestry,  Virginia  Polytechnic  Institute  &  St.  Univ.,  Blacksburg,  VA,  USA. 


667 


Zellner  (1962),  and  Zellner  and  Theil  (1962)  resulted  in  almost  all  methods 
currently  available  for  estimating  the  parameters  in  intercorrelated  (simultaneous) 
systems  of  equations. 

The  objective  of  this  paper  is  to  evaluate  and  compare  joint  versus 
separate  regression  techniques  for  single  tree  growth  and  yield  modeling.  We 
estimate  the  parameters  of  individual  tree  models  for  basal  area  increment,  height 
increment,  and  crown  ratio  using  least  squares  methods  separately  and  jointly  by 
applying  two-stage  (Theil  1953)  and  three-stage  least  squares  (Zellner  and  Theil 
1962)  techniques.  We  specifically  investigate  (1)  the  differences  in  the  estimated 
coefficients  and  (2)  the  correlation  between  the  predictions. 

METHODS 
Independent  Regressions 

We  begin  with  a  system  of  three  individual  tree  growth  equations  for  stand 
conditions  in  Austria:  basal  area  increment,  height  increment,  and  crown  ratio. 
These  equations  were  developed  independently  using  the  same  dataset. 

Basal  area  increment:  After  eliminating  the  qualitative  site  descriptors 
chosen  by  Monserud  and  Sterba  (1996),  which  reduced  the  variance  explained 
only  by  2.6%,  we  are  left  with  the  following  model  for  Yr: 

(1) 

ln(AA4)  =  a  +  b^MP)  +  b2D2  +  b3ln(C)  +  cxBAL  +  c2CCF2  +  d^ELEV7,  +  d^-SL2  +  e, 

with  ABA  the  5-year  basal  area  increment  (outside  bark),  D  the  diameter  at  breast 
height  (1.3  m)  in  cm,  C=(1/CR)-1  where  CR  is  the  crown  ratio,  BAL  the  basal 
area  (m2/ha)  of  trees  larger  in  diameter  than  the  subject  tree,  CCF  the  crown 
competition  factor  of  Krajicek  et  al.  (1961),  ELEV  the  elevation  in  hectometers, 
and  SL  the  tangent  of  the  slope  angel  (%/100). 

Height  increment:  Hasenauer  and  Monserud  (1996a)  used  a  similar 
formulation  to  predict  5-year  height  increment  AH,  where  H  is  the  tree  height, 
and  all  other  parameters  as  previously  defined.  The  second  equation  (Y2)  in  the 
system  is: 

ln(AJ7)  =  a  +  bx'MP)  +  b2H2  +  Jyhi(C)  +  cxBAL  +  c2CCF  +  d^ELEV2  +  d2SL  +  a,  (2) 

Crown  ratio:  To  ensure  that  the  predictions  of  crown  ratio  (defined  as  the 
crown  length  divided  by  tree  height)  are  bounded  between  0  and  1,  Hasenauer 
and  Monserud  (1996b)  chose  a  logistic  function.  After  linearizing  the  logistic  and 
rearranging,  we  are  left  with  the  following  logarithmic  transformation  of  crown 
ratio  (Y3=ln(Q): 

(3) 

IdCQ  =  a  +  blHID+b2'H+b)'D2+cxBAL+c2\n(<CCF)+dl'ELEV+^^  +■  % 

where  C=(1/CR)-1,  HID  is  the  height/diameter  ratio  (m/cm),  AZ  is  the  azimuth 
in  radians,  and  all  other  parameters  are  as  previously  defined.  Change  in  crown 
ratio  is  not  available,  because  height  to  crown  base  was  not  remeasured  after  the 
initial  inventory. 


668 


Simultaneous  Equation  Systems 


In  the  system  above,  it  is  likely  that  the  errors  e  in  equations  (1)  -  (3)  are 
intercorrelated  because  they  are  associated  with  various  attributes  of  the  same 
tree.  If  this  is  the  only  common  influence  among  the  three  equations,  then 
Zellner's  (1962)  seemingly  unrelated  regression  (SUR)  procedure  would  be 
appropriate  because  the  equations  are  related  through  contemporaneous 
correlations  in  the  variance-covariance  matrix.  We  write  the  multivariate 
regression  model  as 

Y  =       +  e  (4) 

where  Y  is  a  3n  x  1  vector  of  dependent  (endogenous)  variables,  X  is  the  3n  x 
(P1+P2+P3)  design  matrix,  0  is  the  (Py+Z^+Pi)  x  1  vector  of  coefficients  to  be 
estimated,  and  e  is  the  3«x  1  error  vector.  The  errors  e  have  fixed  mean  E[e]  =0 
and  variance 


(5) 


where  W^ofl  are  the  main  diagonal  variances  and  Wg=Ogl  are  the  covariances, 
with  /  the  ^-dimensional  identity  matrix. 

The  ordinary  least  squares  (OLS)  estimator  of  p  =(pJ,P2,P3)  ls 

b  =  (x'x)1  x'y  <6> 

with  variance 

cov(b)  =  (x'x)-1  (x'Qx)  (x'x)1  C7) 

In  situations  where      <¥  0  and  a(>  +  o2  for  some  constant  a2,  then  b  is 
inefficient  and  the  usual  estimator  of  (7),  namely 

cav(b)  =  (x'x)1  a2  (8) 

is  biased.  Zellners'  (1962)  seemingly  unrelated  regression  procedure  is  a  form  of 

feasible  generalized  least  squares  (Judge  et  al.  1980)  in  which  Q  is  estimated  by 

Q  ,  where  Q  has  the  same  form  as  (5)  but  with  moment  estimators,  say  o,  and 

o~ ,  in  place  of  the  unknown  parameters  o]  and  .  With  Q  ,  an  asymptotically 
efficient  estimator  of  (3  is 


W\2 

Fie]  =  E[ee']  = 

W2\ 

w2 

W23 

=  0 

W32 

P  =  (x'Q  x) 

The  variance  of  p  is  estimated  by 


/A-l„\-l  X'Q  1 


(9) 


fc 

^11 

^13 

E  =  cov(P)  =  (x'Qx)1  = 

c 

C22 

a  / 

c 

c 

^33  ) 

(10) 


669 


For  our  set  of  single-tree  equations  there  are/?i+/?2+/?i=8+8  +  10=26  estimated 
parameters,  plus  the  main  diagonal  partition  of  the  variance-covariance  matrix  for 

each  model  (CVC2,C3).  The  remaining  off  diagonal  partitions  (C12,C13,C23) 
contain  the  cross-equation  covariances  between  each  estimated  parameter. 

If  Xj =X2=X3  then  SUR  is  identical  to  OLS,  even  when  the  cross-equation 
covariances  are  nonzero.  Otherwise,  SUR  provides  an  asymptotically  more 

precise  estimator  of  p  than  OLS,  and  the  comparative  efficiency  of  (3  increases 
with  increasing  cross  equation  correlation  and  with  increasing  dissimilarity  among 
the  regression  matrices  Xls  X2  and  X3. 

The  use  of  ln(C)  =  Y3  in  the  models  for  In(BAI)  and  In  (AH),  eqns.  (1)  and 
(2)  respectively,  precludes  the  straightforward  use  of  SUR  as  outlined.  Whenever 
a  response  or  endogenous  variable  from  one  model  appears  as  a  regressor  variable 

in  another  model,  both  the  OLS  and  SUR  estimator  of  p  will  be  biased  and 
inconsistent  (Judge  et  al.  1980).  In  this  case  two  stage  least  squares  or  2SLS  (see 
Judge  et  al.  1980)  can  be  used  to  provide  consistent  and  asymptotically  unbiased 

estimates  of  p . 

An  even  more  efficient  estimator  was  invented  by  Zellner  and  Theil 
(1962):  three-stage  least  square  (3SLS).  By  combining  Thiel's  (1953)  2SLS 
procedure  with  Zellner' s  (1962)  SUR  procedure,  the  resulting  3SLS  estimators 

of  p  are  consistent  and  asymptotically  more  efficient  than  2SLS.  We  denote  the 
3SLS  estimates  of  p  as  P  and  its  estimated  variance  as 

S  =  cov(P) 

It  is  convenient  to  partition  E  as 


S  = 


( r    c    c  \ 

^11    ^12  ^13 


^•12   ^22  ^23 
V  C\3  ^33  J 


(12) 


where    c    is  the  pj  x  pk  estimated  covariance  matrix  of  p.  and  $k .  For  details 

concerning  the  3SLS  estimator  in  a  forestry  context  see  Murphy  and  Beltz  (1981), 
Borders  and  Bailey  (1986),  and  Gregoire  (1987). 

In  an  individual  tree  model,  estimates  of  £[5^],  E[Y2],  and  E[Y3]  are 

needed  for  each  tree.  Let  x[  denote  the  pL  x  1  row  vector  of  covariate  values  for 
a  particular  tree  for  which  the  estimated  coefficients  will  be  applied,  such  that 

*  -  4  h  (13) 

serve  as  an  estimate  of  E\y19xl].  Let  y2  and  y3  be  similarly  defined. 


670 


The  distributional  properties  of  the  random  errors  e  in  conjunction  with 
the  estimator  of  p.  determine  the  statistical  properties  of  p.  and  y .  The  scalar 

covariance  between  random  variables  y{  and  y2  is  xlCl2x2/,  where  q  is  the  p! 
x  p2  covariance  matrix  of      and  p2>  Given  an  estimate  of  C12,  say,  q    as  in 

(12),  then  the  covariance  between  y  and  y  can  be  estimated  by  xlCl2x2/' 
Therefore,  the  correlation  between  yx  and  y2  can  be  expressed  as 


r 


12 


The  correlation  r13  between  yx  and  y3 ,  and  the  correlation  r23  between  y2  andy3 
can  be  estimated  similarly. 

DATA 

Data  were  obtained  from  the  Austrian  National  Forest  Inventory 
(Forstliche  Bundesversuchsanstalt  1981),  a  systematic  hidden  permanent  sample 
plot  design  over  the  whole  of  Austria,  with  a  5-yr  remeasurement  interval.  In  a 
given  year  a  fifth  of  the  plots  are  remeasured,  ensuring  a  representative  sample 
of  all  Austrian  forests  each  year.  The  total  inventory  comprises  22,000  permanent 
plots.  We  restricted  ourselves  to  the  4,135  forested  plots  containing  remeasured 
Norway  spruce  (Picea  abies),  not  crossed  by  roads,  and  in  a  single  ownership. 

Permanent  sample  plots  were  established  from  1981  to  1985.  Trees  with 
a  diameter  at  breast  height  (DBH,  1.3  m)  larger  than  10.4  cm  were  selected  by 
angle  count  sampling  using  a  BAF  of  4  m2/ha.  Trees  with  a  DBH  between  5  and 
10.4  cm  were  measured  within  a  circle  of  2.6  m  radius  located  at  plot  center; 
smaller  trees  were  not  recorded.  At  plot  establishment,  the  following  data  were 
recorded  for  every  sample  tree:  species,  DBH  to  the  nearest  mm,  and  distance 
and  azimuth  from  plot  center.  Total  height  and  height  to  the  crown  base  were 
measured  to  the  nearest  decimeter  on  every  fifth  tree.  Plot  descriptors  were 
evaluated  within  a  circle  of  300  m2.  Elevation  is  measured  to  the  nearest  100  m, 
slope  is  measured  to  the  nearest  10%,  and  aspect  to  the  nearest  45°.  Additional 
site  descriptors  were  measured  but  not  used  in  this  comparison  study. 

Plots  were  remeasured  from  1986  to  1990,  5  years  after  establishment.  In 
the  remeasurement,  height  and  diameter  were  measured  on  the  same  trees,  but  not 
height  to  crown  base.  In  summary,  observations  of  7,797  Norway  spruce  trees 
from  4,135  different  permanent  sample  plots  are  available  throughout  Austria. 

RESULTS 

Coefficient  estimates  by  OLS  vs.  3SLS 

Using  the  SYSLIN  procedure  in  the  Econometrics-Time  Series  module  of 


671 


SAS  (SAS  Institute  1988),  the  parameters  in  equations  (1)  to  (3)  were  first 
estimated  independently  by  applying  ordinary  least  squares  (OLS)  and  then  jointly 
by  using  2SLS  and  3SLS. 

Attention  was  immediately  focused  on  the  ln(C)  and  CCF  terms  in  the 
height  increment  model  (2).  These  two  terms  were  both  significant  (a =0.05)  with 
OLS.  With  2SLS,  the  ln(C)  term  remained  significant  but  the  CCF  term  became 
strongly  non-significant.  With  3SLS,  both  the  ln(C)  and  CCF  terms  were  non- 
significant. Thus  the  height  increment  model  (2)  was  reduced  from  8  parameters 
to  6.  The  size  of  the  matrix  partitions  in  the  variance-covariance  matrix  in  (12) 

is  reduced  accordingly.  Now  C22  is  a  6x6  variance-covariance  matrix  for  Y2,  and 

the  off-diagonal  partitions  c[2  and  c'23  are  8x6  and  6x10  matrices  containing  the 

cross-equation  covariances  between  the  estimated  parameters,  respectively. 
Because  there  are  now  /7i+/?2+/7i=8+6  +  10=24  estimated  parameters  in  the 
three  growth  equations,  there  are  576  variance-covariance  elements  in  (12),  a 
reduction  of  100  elements. 


Correlation  between  predictions 

To  investigate  the  strength  of  the  interrelationships  among  our  system  of 
equations  and  evaluate  their  importance,  the  correlations  between  the  predictions 
of  In  (ABA),  In  (AH)  and  ln(C)  (eqns.  (1),  (2),  and  (3))  are  calculated  for  each 
observation  according  to  equation  (14). 

Fig.  1  displays  the  cross-equation  correlations  Z  between  each  pair  of 

ij 

predictions  vs.  the  respective  predicted  variables  Y..  In  Fig.  la  and  lb,  f  is 
massed  around  0.23,  with  a  maximum  correlation  of  0.33  between  the  predictions 
for  In(BAI)  ( j  )  and  In  (AH)  (J  ).  This  indicates  that  the  first  two  equations  in  the 

system  are  fairly  interdependent.  The  correlations      between  predictions  from 

the  first  (J  )  and  third  equations  (y3)  are  weakly  correlated,  with  a  mean  of 

-0.07,  and  extrema  at  0.05  and  -0.11.  The  correlations  f   between  predictions 

from  the  second  (y  )  and  third  equations  (y  )  have  a  mean  of  zero  (0.01),  but 

are  more  dispersed  than  those  for  f  ,  ranging  from  0.26  to  -0. 10  with  a  standard 

deviation  of  0.08.  These  correlations  between  the  basal  area  and  height 
increment  models  and  between  the  height  increment  and  crown  models  confirm 
our  assumption  of  cross-equation  correlations. 


672 


035 
025 
015 
0.05 


S3 


I* 


-0.05 
-015 

0.35 
0.25 
015 
0.05 
-0.05 
-015 


z* 


025 


025 


015 


0.05 


-0.05 


-015  i 


—I  1  1  1  1  1  1  1  1  1  r 

2        3        4        5       6  7 


4       5  6 


-2  -1 


0 


9 


Z^ 


035 
025 
0J5 
0.05 


-0.05 


-015  i 


-2  -1 


0351 


0.25 


S3 


015 


Z* 


0.05 
-0.05] 
-015 


-4     -3     -2     -1  0 


025 


025 


0J5 


Z* 


0.05 
-0.05H 
-015 


-4     -3     -2     -1  0 


Kig.  I:  The  cross  equation  correlation  f  between  each  pair  of  ^  predictions;^ 

indicates  the  predicted  In  (ABA),  y  the  predicted  In  (AH),  and  ^  the  predicted 
ln((//CR)-h. 


673 


DISCUSSION 


Because  individual  tree  forest  growth  models  are  based  on  multivariate 
attributes  observed  on  the  same  individuals  (e.g.  basal  area  increment,  height 
increment,  crown  ratio),  the  resulting  set  of  growth  equations  can  be  considered 
a  simultaneous  system.  Therefore,  joint  regression  techniques  should  be 
considered  for  simultaneously  determining  the  parameter  estimates.  If  endogenous 
variables  do  not  appear  on  the  right  hand  side  (RHS),  the  seemingly  unrelated 
regression  (Zellner  1962)  will  still  improve  the  efficiency  of  the  parameter 
estimates.  If  endogenous  variables  are  used  as  predictor  variables  then  multi-stage 
estimation  techniques  (2SLS  or  3SLS)  are  necessary  to  obtain  parameter  estimates 
that  are  consistent;  such  estimates  will  also  be  asymptotically  efficient.  Ignoring 
the  simultaneous  nature  of  the  system  by  separately  applying  OLS  to  each  model 
can  result  in  estimates  that  are  biased  and  inconsistent. 

Using  3SLS  with  our  system  of  equations  allowed  the  detection  and 
deletion  of  two  non-significant  terms  (CCF  and  ln(C)  in  eq.  (2))  that  OLS  had 
determined  as  significant.  This  led  to  a  simplification  of  the  simultaneous 
structure  of  our  system  because  ln(C)  was  originally  considered  as  an  endogenous 
variable  on  the  RHS. 

One  of  the  advantages  of  joint  regression  techniques  is  that  both  the 
correlation  among  the  predictor  variables  within  a  certain  equation  and  across  all 
equations  are  available.  Fig.  1  indicates  that  cross-equation  may  change  their 
degree  of  dependency  in  the  system  and  that  these  correlations  can  be  used  in  a 
simulator  to  better  predict  stand  and  tree  dynamics.  Predictions  between  the  first 
two  growth  equations  (basal  area  and  height  increment)  are  fairly  well  correlated, 
while  correlations  between  predictions  from  the  first  and  third  (height  increment 
and  crown  ratio)  are  rather  weak. 

A  C  KNOWLEDG  EMENTS 

This  research  was  conducted  when  Hubert  Hasenauer  was  a  Visiting  Scientist  at  Virginia 
Tech  Department  of  Forestry  in  Blacksburg,  and  at  the  Intermountain  Research  Station's  Forestry 
Sciences  Laboratory  in  Moscow,  Idaho.  Hasenauer  was  working  on  a  Sehrddinger  research  grant 
from  the  Austrian  Science  Foundation.  We  are  grateful  to  Karl  Schieler  and  Klemens  Schadauer 
or  the  Federal  Forest  Research  Center  in  Vienna  tor  making  the  Forest  Inventory  data  available. 


674 


LITERATURE 


Aitken,  A.C.  1934-35.  On  the  least-squares  and  linear  combination  of  observations.  Proceedings 

of  the  Royal  Society  of  Edinburgh.  55:  42-48. 
Borders,  B.E.,  and  R.L.  Bailey.  1986.  A  compatible  system  of  growth  and  yield  equations  for 

slash  pine  fitted  with  restricted  tree  stage  least  squares.  For.   Sci.  32:  185-201. 
Burkhart,  H.E.,  Farrar,  K.D.,  Amateis,  R.A.,  and  Daniels,  R.F.,  1987.  Simulation  of  individual 

tree  growth  and  stand  development  in  loblolly  pine  plantations  on  cutover,  site-prepared 

areas.  Coll.  For.  and  Wildlife  Resources.  Virg.  Tech.  Inst.,  Blacksburg,  Va.  Publication 

FWS-1-87:  47  pp. 

Dixon,  R.K.,  Meldahl,  R.S.,  Ruark,  G.A.,  and  Warren,  W.G.  1990.  Process  modeling  of  forest 
growth  responses  to  environmental  stress.  Timber  Press,  Portland,  OR.  441  pp. 

Forstliche  Bundesversuchsanstalt,  1981.  Instruktionen  fur  die  Feldarbeit  der  Osterreichischen 
Forstinventur  1981-1985.  Forstliche  Bundesversuchsanstalt,  Wien,  172  pp. 

Gregoire,  T.G.  1987.  Generalized  error  structure  for  forestry  yield  models.  For.  Sci.  33:  423- 
444. 

Haavelmo.  T.  1943.  The  statistical  implications  of  a  system  of  simultaneous  equations. 
Econometrica  II:  1-12. 

Hasenauer,  H.  1994.  Ein  Einzelbaumwachstumssimulator  fur  ungleichaltrige  Fichten-Kiefern-  und 

Buchen-Fichtenmischbestande.   Forstl.  Schriftenreihe,  Univ.  f.   Bodenkultur,  Wien. 

Osterr.  Ges.  f.  Waldokosystemforschung  und  experimentelle  Baumforschung  an  der 

Univ.  f.  Bodenkultur.  Band  8:  152  pp. 
Hasenauer.  H..  and  R.A.  Monserud.  1996a.  Biased  statistics  from  height  increment  predictions. 

(submitted  to  Ecol.  Modeling). 
Hasenauer.  H..  and  R.A.  Monserud.  1996b.  A  crown  ratio  model  for  Austrian  forests.  Forest 

Ecology  and  Management  (in  press). 
Judge.  G.G..  W.E.  Griffiths.  R.C.  Hill,  and  T.C.  Lee.    1980.  The  theory  and  practice  of 

econometrics.  John  Wiley  &  Sons,  Inc.,  New  York.  810  pp. 
Krajieek.  .I.E..  K.A.  Brinkman.  and  S.F.  Gingrich.  1961.  Crown  competition:  A  measure  of 

density.  For.  Sci.  7:  35-42. 
Monserud.  R.A.  1975.  Methodology  for  simulating  Wisconsin  northern  hardwood  stand  dynamics. 

Ph.D.  Thesis,  Univ.  of  Wisconsin,  Madison.  156  pp. 
Monserud,  R.A.,  and  H.  Sterba.  1996.  A  basal  area  increment  model  for  individual  trees  growing 

in  even-  and  uneven-aged  forest  stands  in  Austria.  For.  Ecol. Manage,  (in  press). 
Murphy.  P.  A  .  and  R.C.  Beltz.  1 98 1 .  Growth  and  yield  of  shortleaf  pine  in  the  West  Gulf  region. 

South.  Forest  Exp.  Station.  New  Orleans,  LA.  USDA  For. Serv. Res. Paper  SO- 169:  15 

pp. 

SAS  Institute.  1988.  SAS/ETS  user's  guide,  version  6.  Cary.  NC.  560  pp. 

Theil.  H.  1953.  Repeated  least-squares  applied  to  complete  equation  systems.  The  Hague:  The 

central  planing  bureau.  The  Netherlands. 
Wykoff,  W.R..  Crookston.  N.L.  and  Stage,  A.R.,  1982.  User's  Guide  to  the  Stand  Prognosis 

Model.  USDA  For.  Serv.  GTR  INT-133.  1  12  pp. 
Zellner,  A.  1962.  An  Efficient  method  of  estimating  seemingly  unrelated  regressions  and  tests  for 

aggregation  bias.  Journal  of  the  American  Stat.  Association.  57:  348-368. 
Zellner.    A.,   and  Theil,    H.    1962.   Three-stage   least  squares:   simultaneous  estimation  of 

simultaneous  equations.  Econometrica  30:  54-78. 


675 


Genetics,  geographies,  and  prairie  dogs: 
error  and  accuracy  in  a  validated 
spatially-explicit  dispersal  model. 

Gillian  Bowser1 


Abstract.—  Genetic  theory  predicts  that  populations  will  differentiate 
based  on  the  geographic  distance  between  them  unless  some 
environmental  feature  creates  a  barrier  to  dispersal.  However,  those 
barriers  are  often  inferred  from  landscape  features,  and  rarely  directly 
tested  for  their  ability  to  predict  genetic  differentiation  as  expressed  by 
genetic  distances.  There  is  a  need  to  develop  spatially  explicit  models 
validated  by  population  characteristics,  such  as  genetics,  to  predict 
population  sub-division— an  important  step  in  the  conservation  of  a 
species.  Prairie  dogs  (Cynomys  ludoviciartus),  are  colonial  fossorial 
rodents  that  live  in  discrete  populations  which  have  distinctive  signatures 
in  remotely-sensed  images,  providing  a  mechanism  to  combine  spatial  and 
genetic  databases  into  a  single  model.  However,  this  combination  of 
spatial  and  genetic  data  complicates  spatial  error  as  spatial  genetic 
parameters,  measured  from  protein  differcntion,  have  associated  error 
around  distance  algorithms;  while  spatial  databases,  such  as  elevation, 
hydrology,  and  transportation,  have  error  associated  with  placement.  The 
implications  of  testing  spatial  models  on  small  organisms  at  the  landscape 
level  are  discussed 


INTRODUCTION 

Ecological  modeling  of  animal  populations  with  remotely-sensed  data  is  a  relatively 
new  field  for  conservation  biologists  (Johnson  1990).  How  animals  move  among 
populations  and  how  different  environmental  features  affect  population  exchange  and 
stability  are  fundamental  questions  for  the  conservation  of  patchy  fragmented  habitats 
(Johnson  and  Naiman  1987,  Jensen  et  al.  1990).  Given  the  current  availability  of 
remotely-sensed  databases  along  with  the  statistical  tools  to  analyze  those  databases,  there 
is  also  an  increasing  need  to  address  the  accuracy  and  resolution  of  these  landscape 
databases  when  used  at  the  scale  of  an  individual  animal  (Costanza  and  Maxwell  1994). 

Modeling  population  establishment  and  movement  on  a  landscape  level  requires 
using  some  remotely-sensed  data  (such  as  vegetation,  elevation,  or  the  like)  to  imply 
where  the  animal  populations  may  occur  (Scott  et  al.  1993).  However,  the  digital 
databases  expressing  the  environmental  parameters  contain  sources  of  error  associated 


Natural  Resource  Specialist,  Rocky  Mountain  System  Support  Office,  National  Park  Service,  Denver, 
Colorado 

676 


with  the  scale,  resolution  and  projection  of  the  source  map  that  can  have  important 
implications  when  applied  to  ecological  databases  (Meenetemeyer  and  Box  1987,  Lee  et 
al.  1992,  Kemp  1993,  Bolstad  and  Stowe  1994,  Hodgson  1995).  In  addition,  the  unit  of 
scale  of  the  landscape  will  change  with  the  different  ecological  processes  being  examined 
and  will  vary  with  different  organisms  as  to  the  correct  scale  for  expressing  spatial 
movements  (Hunsacker  et  al.  1993  Cale  and  Hobbs  1994).  At  a  certain  point,  the  error 
associated  with  the  scale  of  the  underlying  database  can  overwhelm  any  potential  to  model 
populations  at  finer  resolutions  (Brown  and  Bara  1994,  Berry  1995).  This  point-of-no- 
return  can  be  expressed  as  an  error  term  or  through  error  management  to  evaluate  model 
performance  (Hunter  and  Goodchild  1995). 

Here  I  discuss  the  implications  of  scale  and  resolution  on  constructing  a  spatially 
explicit  model  of  animal  dispersal,  and  how  the  validation  of  this  model  is  limited  by 
discrepencies  between  the  ecological  scale  (how  far  an  individual  animal  can  move  on  a 
landscape)  and  the  database  scale  (what  are  the  limitations  of  resolution  and  projection  on 
the  database  accuracy  for  modeling  the  movements  of  small  mammals). 

BACKGROUND 

Dispersal  among  populations  is  essential  to  maintain  genetic  homogeneity  within  a 
population  thus  preserving  population  health  and  longetivity  (Green  1994).  As 
populations  become  isolated  through  environmental  or  behavioral  barriers,  genetic 
divergence  occurs  leading  eventually  towards  speciation.  The  progress  towards  speciation 
can  be  estimated  on  a  shorter  time  scale  by  calculating  the  dispersal  rates  per  generation 
and  creating  a  genetic  distance  (based  on  exchange  as  mirrored  by  genetic  differentiation-- 
Nei  1972)  that  can  validate  dispersal  distances  (a  rate  of  exchange  based  on  physical 
separation). 

To  examine  dispersal  distances  across  realistic  landscapes  as  a  measure  of  population 
conservation,  each  element  of  a  landscape  (elevation,  vegetation,  transportation, 
hydrology)  can  be  viewed  as  an  independent  treatments  (single  model  runs)  to  be  tested 
against  the  predicted  dispersal  distances  based  on  genetic  differentiation  (as  a  measure  of 
population  persistence.  The  effects  of  environment  elements  (such  as  roads  or 
developments)  on  dispersing  animals  can  be  directly  by  calculating  population  parameters 
(genetic  differentiation)  against  an  estimate  of  the  exchange  rate  (a  dispersal  distance)  that 
creates  a  spatially-explicit  dispersal  model  applicable  for  population  conservation  efforts. 

Model  Parameters 

Three  aspects  of  ecological  model  are  directly  affected  by  the  scale  and  resolution  of 
the  spatial  databases  themselves:  the  size  of  the  area  covered  by  the  model;  the  size  of  the 
species  the  model  is  designed  to  emulate;  and  the  variablity  in  the  landscape. 

How  big  is  the  study  area? 

The  first  step  in  model  construction  is  to  define  the  size  of  the  patch  addressed  by 
the  model.  If  the  patch  is  a  preserve  or  politically  defined  boundary,  there  may  be 
constraints  defined  by  the  boundary  of  that  patch  which  should  be  included  in  any 


677 


modeling  efforts.  In  addition,  the  size  of  the  total  area  of  consideration  can  determine  the 
resolution  of  the  databases  feasible  given  hardware  and  software  limitations.  Highly 
mobile  insects,  for  example,  would  required  extremely  high  resolution  data  over  large 
areas  that  is  beyond  the  storage  capacity  of  many  computers. 

How  large  is  the  organism? 

The  size  of  the  species  will  dictate  the  required  precision  of  any  underlying  map 
layers  (Brown  and  Bara  1994).  Smaller  species  may  be  more  sensitive  to  the  resolution  of 
a  spatial  data  than  larger  species  and  thus  a  model  based  on  a  coarser  resolution  data 
could  miss-  or  over-represent  landscape  features  and  their  impact  on  animal  movement. 
The  physical  size  of  the  species  will  also  dictate  how  rapidly  dispersal  declines  with 
geographic  distance,  in  which  case  modeling  population  placement  may  be  more  valid  than 
individual  movement  (even  though  this  is  fundamentally  a  different  question)  and  the 
resolution  of  digital  layers  maybe  unrealistic  for  that  species  (Cale  and  Hobbs  1994). 

How  variable  is  the  study  area? 

Habitat  variablity,  especially  as  it  affects  small  animals,  can  be  under-represented  by 
databases  with  coarse  resolution.  For  example,  small  erosive  features,  long  narrow  gullies 
or  streambeds,  may  not  be  visible  to  a  database  with  a  minimum  pixel  size  of  30meters.  If 
those  gullies  are  barriers  to  the  movement  of  small  animals,  the  database  may  not  be  able 
to  detect  those  barriers  and  could  mis-represent  animal  movement. 

STUDY  SITE 

Black-tailed  prairie  dogs  (Cynomys  ludovicianus)  are  colonial  fossorial  rodents  of 
the  Great  Plains  of  North  America.  The  prairie  dog  is  considered  a  keystone  species  for 
the  North  American  plains  because  they  form  large  colonies  and  actively  modify  the 
environment  within  their  colonies,  creating  unique  ecosystems.  The  colonies  with  their 
distinct  vegetation  are  distinct  in  aerial  photography  (Schenbeck  and  Myre  1990)  and  thus 
are  ideal  to  demonstrate  spatially-explicit  dispersal  models.  Badlands  National  Park, 
South  Dakota  has  a  number  of  large  prairie  dog  colonies  scattered  throughout  the  park. 
As  the  park  has  two  distinct  lobes  with  colonies  on  either  end,  predicted  dispersal  patterns 
among  colonies  are  colonies  relatively  close  on  one  side  of  the  park  would  be  genetically 
more  similar  than  those  on  opposite  (figure  1).  These  predictions  (euclidean  distances 
describe  population  dispersal  patterns)  will  serve  to  validate  model  predictions  based  on 
the  dispersal  and  genetic  distances. 


678 


Figure  1.  Cartoon  of  the  distribution  of  major  prairie  dog  populations  in  Badlands  National  Park, 
South  Dakota.  Population  abbreviations  are:  Baysinger  (BA),  Burns  Basin  (BU),  Bigfoot  (BI),  Haybutte 
(HA),  Kocher  Flats  (KO),  Tyree  Basin  (TY),  Tyree  South  (TS). 

MODEL  CONSTRUCTION 

Environmental  layers  such  as  Digital  Elevation  Model  (DEM- 1:24, 000),  Digital  Line 
Graph. (DLG-- 1:25 0,000),  aerial  photography  (black  and  white  and  color  infra-red: 
1:24000),  Soil  Conservation  Service  (SCS  DLG-- 1:250, 000),  vegetation  (National  Park 
Service  (NPS)  database- 1:24000)  and  landuse  boundaries  (NPS  database- 1:250,000) 
were  entered  into  a  GRASS4.1  GIS  database  on  a  Sun  Sparkstation.  All  data  were 
rectified  with  GRASS  program  modules  for  geo-referencing  (Clark66)  and  spheroid 
corrections  (NAD27).  Additional  corrections  for  resolution  were  done  using  C-  and 
Bourne-shell  scripts  and  GRASS  resampling.  All  model  runs  were  generated  by  these 
scripts  and  were  run  in  the  UMX  environment,  calling  the  GRASS  modules  as  needed. 

A  model  run  consisted  of  an  individual  GRASS  surface  (such  as  roads)  redefined 
based  on  user-assigned  weights  meant  to  symbolize  the  ease  or  difficulty  for  prairie  dogs 
to  cross  a  particular  environmental  feature.  Those  weights  are  then  expressed  as  a  cost 
surface  (the  cummulative  weights  for  moving  from  a  particular  point  to  any  other  point  on 
the  surface)  which  is  then  used  to  generate  a  dispersal  path  (figure  2).  Dispersal  paths  are 
calculated  using  alogorithms  from  hydrology  as  water  seeks  the  lowest  path  down  a 
drainage.  The  route  water  would  follow  then  represents  the  least-cost  path  that  an  animal 
would  follow  over  an  ecologically-defined  surface  (figure  3). 

Dispersal  paths  were  validated  by  using  the  genetic  differentiation  of  proteins  (as 
determined  by  the  horizontal  gel-electrophoresis  of  polymorphic  enzymes)  to  create  a 
genetic  distance  analogous  to  the  physical  separation  (see  Bowser  1996  for  methods). 
Genetic  distance  is  a  metric  measure  of  the  genetic  differentiation  of  proteins  among 

679 


population  samples,  and  can  be  used  as  an  independent  measure  to  test  for  linear 
associations  with  the  paths  generated  by  the  dispersal  models  (Nei  1972,  Slatkin  1985). 
Genetic  distances  were  calculated  using  BIOSYS-1  (Swofford  and  Selander  1989)  and 
Weirs  (1990)  jack-knifing  procedure. 


mm 

mm 

mm 


ma 


11  

i 

III 

:<■:■:■:■■  ::-A 


111 


IL 

mm 

::::::::::::::::::: 

mm*; 


Visual  summary  of  cost  path  generation  using 
several  transforming  surfaces 


Surface  X 


Surface  Y 


Cost  surface 


Transforming 
surfaces 


Least-cost 
path 


mm 

mm 

mm 


il 
II 


m 


I 


Figure  2.  Conceptual  diagram  of  the  cost  path  generation  using  surfaces  consisting  of  single 
environmental  feature  to  create  composite  surfaces.  Populations  are  represented  by  T  and  'j'  and  paths 
are  calculated  from  the  centroids  of  i  and  j.  Different  polygons  indicate  ecological  features  on  the 
transforming  surface  that  pose  a  barrier  to  dispersal. 


All  statistical  tests  were  done  using  Splus3.3  in  the  UNIX  environment  after 
corrections  for  normality  and  checks  for  autocorrelation.  Each  environmental  surface  was 
seen  as  an  independent  model  consisting  of  a  single  parameter  such  as  roads,  streams,  etc. 
Composite  layers  were  created  within  GIS  from  combinations  of  single  layers  which  were 
then  rescaled  to  account  for  changes  in  resolution  and  tested  as  a  single  independent 
surface 


RESULTS 


Dispersal  paths  calculated  for  prairie  dog  populations  in  BADL  were  significantly 
different  than  the  Euclidean  distances  among  the  populations  (table  1).  All  of  the 
environmental  surfaces  created  longer  paths  than  the  Euclidean  distances  with  the 
exception  of  vegetation  which  was  not  significantly  different.  Loglinear  regression  on 
splined  distances  showed  that  there  were  weak  predictive  relationships  (p=0.05,  r2=.  14) 
for  single  models  between  the  genetic  distances  and  the  dispersal  distances  (figure  4).  The 
best  fitted  model  using  step-wise  regression  was  roads,  streams,  and  park  boundary 


680 


(p=0.05, 1^=0.33).  Composite  models  were  generally  better  at  predicting  the  genetic 
distances  than  single  models  using  Principal  Component  Analysis;  composite  models 
contributed  the  most  to  the  first  PCA  axis  which  explained  86%  of  the  variance  in  genetic 
distances. 


Figure  3.  Cost  surface  for  moving  from  one  location  (the  pit)  to  any  other  point  on  the  surface. 
Cost  re  fers  to  the  cummulative  sum  of  pixel  weights  from  a  designated  starting  point  to  any  other  point 
on  the  surface. 


Table  1.  ANOVA  on  single  surface  models  as  independent  treatments  against  Euclidean  distance 
among  populations  of  prairie  dogs  in  Badlands  National  Park.  


Model  df  f  p  sign 


roads  1 

64.55 

0.00 

** 

streams  1 

75332.6 

0.00 

** 

slope  1 

159.97 

0.00 

** 

elevation  1 

286.23 

0.00 

** 

vegetation  1 

1.20 

0.28 

NS 

DISCUSSION 

The  validated  dispersal  models  demonstrated  here  showed  an  ability  to  predict 
another  population  parameter,  genetic  distance,  which  is  an  important  measure  of  the 
longetivity,  exchange,  and  isolation  of  populations.  As  an  exploratory  model,  the  dispersal 
model  was  able  to  demonstrate  which  environmental  layer  appear  to  best  explain  the 
genetic  patterns  observed,  as  well  as  provide  some  insights  as  to  which  environmental 
features  may  have  little  effect.  Despite  the  presence  of  the  badlands  wall  and  other 
erosion  features  within  BADL,  elevation  was  not  a  strong  predictor  of  genetic  distance. 

681 


In  contrast,  streams  and  roads  together  as  a  composite  model  was  best  able  to  explain  the 
genetic  patterns  over  any  single  environmental  feature. 


Figure  3.  Regression  lines  for  each  single  surface  tested  using  genetic  distances  as  the  dependent 
variable.  Graphs  labels  from  the  left:  roads,  streams,  slope,  vegetation,  elevation,  and  park  boundary. 

The  Influence  of  Error 

There  are  two  sources  of  error  discussed  here:  error  associated  with  the  placement 
of  populations;  and  error  associated  with  treatment  affects  during  the  model  validation. 

Population  placement 

The  location  of  populations  on  the  landscape  for  a  dispersal  model  is  critical  as  the 
seed  point  for  constructing  paths  on  the  cost  surface.  If  a  centroid  position  for  a 
population  polygon  is  used  as  the  starting  point,  the  error  associated  with  that  centroid 
only  becomes  important  when  referenced  to  ground  measurements.  Each  centroid  has  a 
sphere  of  error  round  it  that  represents  the  corrections  due  to  population  placement 
(digitizing  and  interpretation  error),  RMS  associated  with  the  photograph  rectification, 
along  with  any  error  associated  with 

Treatment  error 

Treatment  error  builds  on  existing  population  placement  error  within  the  framework 
of  a  statistical  test.  Location  error  can  be  analogous  to  the  variance  around  each  point 
within  a  treatment.  Testing  the  variance  among  treatments  (assuming  each  treatment  is  a 
different  spatial  layer),  changes  the  total  error  incorporated  within  the  statistical  model 
would  be  a  combination  of  the  variance  among  and  within  treatments—the  first  due  to 
population  placement  and  the  second  due  to  the  underlying  precision  and  resolution  of  the 
original  environmental  database.  The  combined  errors  reduce  the  power  of  a  variance  test 
to  explain  the  "background  noise"  resulting  in  low  multiple  r  squared  values. 

CONCLUSIONS 


Error,  scale,  and  resolution  can  limit  the  ability  of  GIS  to  accurately  model  animal 

682 


movements  across  complex  landscapes.  When  a  model  is  validated  by  a  population 
measure  (such  as  genetic  distances)  the  error  incorporated  into  spatial  models  from  digital 
databases  and  the  assumptions  of  accuracy  underlying  those  databases  are  elucidated.  In 
the  dispersal  model,  the  inability  to  explain  more  than  33%  of  the  model  variance  in  the 
regression  models  yet  still  have  a  demonstratable  relationship  (and  reproduce  patterns 
expected  by  the  population  parameters)  suggests  that  the  error  in  placement  may  distort 
model  results.  The  take-home  message  is  that  although  the  dispersal  model  did  a  modest 
job  in  re-creating  the  genetic  patterns  and  sorting  through  which  environmental  features 
best  explained  those  patterns,  the  accuracy,  precision,  and  resolution  of  the  underlying 
databases  must  be  discussed.  In  this  model,  the  minimum  resolution  (30  meters)  was 
greater  than  most  of  the  animal's  home  territories  (and  some  colonies).  Realistically,  the 
model  requires  the  animal  to  move  in  jumps  of  at  least  one  meter  up  to  30  meters 
depending  on  the  surface—a  difficult  task  for  an  animal  of  2-10  pounds.  The  fact  that  the 
model  still  distinguished  among  surfaces  and  replicated  genetic  distance  predictions  is 
intriging  and  will  hopefully  generated  further  exploration  as  to  the  effects  of  the  accuracy 
and  error  of  spatial  databases  on  ecological  modeling. 

ACKNOWLEDGEMENTS 

Bette  Loiselle  and  John  Blake,  my  dissertation  advisors  from  the  University  of 
Missouri-St.  Louis,  provided  strong  support  in  the  development  of  this  project  as  a  part  of 
my  dissertation.  Chip  Harvey,  Chris  Theriault,  and  Michelle  Gudorf  of  the  National  Park 
Service  provided  excellent  discussions  on  spatial  error  and  ecological  modeling.  I 
appreciate  the  efforts  of  Peter  Strong,  David  Duran,  and  Bruce  Powell  of  the  National 
Biological  Service  (NBS)  to  keep  my  computer  running.  Ralph  Root  and  Susan  Stitt 
(NBS)  provided  excellent  analytical  support  for  the  spatial  analysis. 

LITERATURE  CITED 

Bolstad,  P.  V.  and  T.  Stowe  1994.  An  evaluation  of  DEM  accuracy:  elevation,  slope  and 

aspect.  Photogrammetric  Engineering  and  Remote  Sensing  60(1 1):  1327-1332. 
Costanza,  R.  and  T.  Maxwell  1994.  Resolution  and  predictability:  an  approach  to  the 

scaling  problem.  Landscape  Ecology  9(1):  47-57. 
Bowser,  G.  1996.  Integrating  ecological  tools  with  remote  sensing:  modeling  animal 

dispersal  on  complex  landscapes.  Conference  proceeding  Third  Annual  Conference  on 

Environmental  Modeling  and  GIS,  Santa  Fe,  New  Mexico,  CDRom. 
Brown,  D.  G.  and  T.  J.  Bara  1994.  Recognition  and  reduction  of  systematic  error  in 

elevation  and  derivative  surfaces  from  7 1/2  minute  DEMs.  Photogrammetric 

Engineering  and  Remote  Sensing  60(2):  189-194 
Goodchild,  M.  F.  1 993 .  The  state  of  GIS  for  environmental  problem  solving  in 

Environmental  Modeling  with  GIS.  M.F.  Goodchild,  B.  O.  Parks  and  L.  T.  Steyaert 

editors.  Oxford  University  Press.  485  pages 
Green.  D.  G.  1994.  Connectivity  and  complexity  in  landscapes  and  ecosystems.  Pacific 

Conservation  Biology  1 : 194-200. 


683 


Hodgson,  M.  E.  1995  What  cell  size  does  the  computed  slope/aspect  represent? 

Photogrammetric  Engineering  and  Remote  Sensing  61(5):  513-517. 
Hunter.  G.  J.  and  M.  F.  Goodchild  1995.  Dealing  with  error  in  spatial  databases:  a  simple 

case  study.  Photogrammetric  Engineering  and  Remote  Sensing  61(5):  529-537. 
Jensen,  J.R.,  S,  Narumalani,  O.  Weatherbee  and  K.  M.  1992.  Predictive  modeling  of 

cattails  and  waterlily  distribution  in  a  South  Carolina  reservoir  using  GIS. 

Photogrammetric  Engineering  and  Remote  Sensing  58:  1561-1568. 
Johnson,  C.  A.  and  R.  J.  Naiman  1990.  The  use  of  geographical  information  systems  to 

analyze  long-term  landscape  alteration  by  beaver.  Landscape  Ecology  4:  5-19. 
Johnson,  L.  B.  1990.  Analyzing  spatial  and  temporal  phenomena  using  geographic 

information  systems:  A  review  of  ecological  applications.  Landscape  Ecology  4:  31-43 
Hunsaker,  C.  T.,  R.  A.  Nisbet,  D.  C.  L.  Lam,  J.  A.  Browder,  W.  L.  Baker,  M.  G.  Turner, 

and  D.  B.  Botkin.  1993.  Spatial  model  of  ecological  systems  and  processes:  the  role 

of  GIS.  in  Environmental  Modeling  with  GIS.  M.F.  Goodchild,  B.  O.  Parks  and  L. 

T.  Steyaert  editors.  Oxford  University  Press.  485  pages 
Kemp,  K.  K.  1993.  Spatial  databases:  sources  and  issues,  in  Environmental  Modeling 

with  GIS.  M.F.  Goodchild,  B.  O.  Parks  and  L.  T.  Steyaert  editors.  Oxford  University 

Press.  485  pages. 

Lee,  L.,  P.  K.  Snyder,  and  P.  F.  Fisher  1992.  Modeling  the  effect  of  data  errors  on  feature 
extraction  from  digital  elevation  models.  Photogrammetric  Engineering  and  Remote 
Sensing  58(10):  1461-1467. 

Schenbeck,  G.  L.  and  R.  J.  Myhre  1986.  Aerial  photography  for  assessment  of  black- 
tailed  prairie  dog  management  on  the  Buffalo  Gap  National  Grassland,  South  Dakota. 
USDA  Report  no  86-7  3400. 

Scott,  J.  M.,  F.  Davis,  B.  Csuti,  R.  Noss,  B.  Butterfield,  C.  Groves,  H.  Anderson,  S. 
Caicco,  F.  D'Erchia,  T.  C.  Edwards  jr,  J.  Ulliman,  and  R.  G.  Wright  1993.  GAP 
analysis:  a  geographic  approach  to  protection  of  biological  diversity.  Wildlife 
Monographs  123. 

Swofford,  D.  L.  and  R.  B.  Selander  1989.  Biosys-1:  A  computer  program  fo  the  analysis 
of  allelic  variation  in  population  genetics  and  biochemical  systematics.  Release  1.7 
Illinois  Natural  History  Association,  Illinois. 

Wickham,  J.  D.  and  D.  J.  Norton  1994.  Mapping  and  analyzing  landscape  patterns. 
Landscape  ecology  9(1):  7-23. 

Wier,  B.  1990.  Genetic  Data  Analysis.  Sinauer  Associates,  New  York 

BIOGRAPHICAL  SKETCH 

Gillian  Bowser  is  a  natural  resource  specialist  with  the  National  Park  Service  in 
Denver,  Colorado.  She  holds  an  MS  from  the  University  of  Vermont  in  Zoology  and  is 
currently  a  PhD  candidate  with  the  University  of  Missouri-St.  Louis.  Gillian  provides 
technical  assistance  and  develops  proposals  in  the  sciences  for  national  parks  in  the  rocky 
mountain  area. 


684 


Effect  of  Uncertainty  in  Mapped 
Biodiversity  Data  on  Optimal 
Conservation  Decisions 


Michael  J.  Conroy1  and  Jennifer  E.  Crocker2 


Abstract-Spatially  referenced  biological  diversity  data  are  increasingly 
being  used  to  assist  with  conservation  decisions.  Examples  include  the 
identification,  selection,  and  design  of  ecological  reserves;  the  types, 
amount,  and  distribution  of  forest  cutting;  and  the  placement  of  corridors 
to  connect  reserves.  These  have  in  common:  (1)  an  implicit  long-term 
objective;  (2)  dynamic  systems  ;  (3)  alternative  possible  decisions  ;  (4) 
alternative  uncertain  outcomes;  (5)  uncertainty  about  both  the  current 
system  state  (sampling  errors),  and  in  predictions  about  future  system 
states  ("model  uncertainty").  Decision-theoretic  methods  and  predictive 
models  can  be  used  to  find  decisions  that  result  in  a  long  term  optimum 
(e.g.,  a  balance  between  species  conservation  and  timber  revenue),  subject 
to  updating  through  current  surveys,  research,  and  other  information.  We 
illustrate  the  approach  for  a  hypothetical  reserve  design  problem  involving 
four  species  with  different  habitat  affinities  in  a  landscape  with  two  habitat 
types,  with  both  demographic  and  statistical  uncertainty.  In  particular,  we 
examine  how  uncertainty  in  predictive  models  (e.g.,  models  relating 
animal  distribution  and  abundance  to  habitat  features)  can  result  in 
suboptimal  decisions,  and  how  research  and  monitoring  can  be  used 
adaptively  to  reduce  uncertainty  and  improve  decision  making. 

INTRODUCTION 

Increasingly,  managers  are  faced  with  making  decisions  about  how  to  manage 
systems  in  which  there  are  multiple,  competing  goals  or  objectives,  and  the  system 
under  management  is  both  complex  and  incompletely  understood.  The  failure  of 
public  decision  makers  to  make  a  "good"  decision  can  have  severe  consequences, 
ranging  from  loss  of  agency  credibility  and  good  will,  to  unnecessary  losses  of 
revenue  or  opportunity,  to  adverse  impacts  on  the  resource,  including  the  loss  of 
species.       Concern  over  loss  of  biological  diversity  has  led  to  efforts  in  the 


Assistant  Leader.  Georgia  Cooperative  Fish  and  Wildlife  Research  Unit,  Athens,  GA 
Graduate  research  assistant.  Institute  of  Ecology,  University  of  Georgia.  Athens.  GA 


685 


cartographic  analyses  of  species  distributions  and  landscape  characteristics  (e.g., 
Burley  1988;  Scott  et  al.  1993)  In  these  efforts,  animal  species  richness  is  either 
observed  directly  or  inferred  from  vegetation  (habitat)  maps  and  models  relating 
abundance  or  presence/absence  to  vegetation  cover  attributes  and  physical  features 
(Scott  et  al.  1993).  The  mapped  data  are  then  used  land-use  decisions  directed  at 
the  conservation  of  biological  diversity,  such  as  reserve  design  (Scott  et  al.  1993). 

Biodiversity  mapping  programs  such  as  Gap  Analysis  frequently  do  not  formally 
incorporate  uncertainty  about  1)  mapping  errors  in  habitat  and  animal  species 
presence  distributions,  and  2)  species-habitat  associations.  They  implicitly  assume 
that  predictions  can  be  made  about  the  future  state  of  biodiversity,  given  alterations 
in  current  habitat  conditions  (e.g.,  through  management).  Here  we  consider  how 
statistical  unreliability  in  spatially  referenced  data  may  affect  the  utility  of  mapped 
species  richness  patterns  for  conservation  decisions  such  as  reserve  selection  and 
design.  We  first  briefly  consider  alternative  approaches  to  obtaining  optimal 
management  decisions  under  the  conditions  in  which  the  system  is  1)  stochastic,  2) 
dynamic,  3)  imperfectly  understood,  and  4)  incompletely  observable.  We  illustrate 
the  general  problem  by  means  of  an  artificial  landscape  containing  2  habitat  types 
and  4  species  (Conroy  and  Noon  in  press,  Crocker  and  Conroy  1995). 

Before  proceeding,  we  note  that  we  are  using  one  specific  method  (stochastic 
dynamic  programming)  for  determining  optimal  policies.  There  are  a  variety  of 
alternative  analytical  approaches  to  this  problem,  capably  reviewed  by  Williams 
(1 989).  These  include  non-dynamic  methods  such  as  linear  programming,  simulation 
methods,  and  combinations  of  simulation  and  optimization.  We  recognize  that  no 
single  approach  is  likely  to  perform  adequately  under  all  circumstances,  but  we 
strongly  feel  that  any  approach  should  adequately  reflect  both  the  dynamic  nature  of 
ecological  systems  and  decision  making,  and  the  role  of  uncertainty  in  both 
observations  and  model  structure.  We  also  agree  with  Walters  (1986)  and  Nichols 
et  al.  (1995)  that  a  formally  adaptive  approach  to  natural  resource  management  will 
in  the  long  run  prove  most  efficacious. 

METHODS 

Elements  of  the  Problem 

We  consider  a  hypothetical  1 00  hectare  landscape  with  three  cover  types,  which 
we  have  termed  forest,  open,  and  developed.  Initially,  the  landscape  is  75%  forested, 
25%  open,  and  0%  developed.  This  landscape  is  initially  occupied  by  four 
hypothetical  species,  two  forest  habitat  specialists  and  two  open  habitat  specialists. 
The  management  objective  expresses  a  tradeoff  between  the  conservation  goal  of 
maintaining  the  species  in  the  landscape  and  the  economic  goal  of  allowing 
development.  At  five  year  intervals,  the  manager  must  decide  what  proportion  of 
each  habitat  type  to  conserve,  and  a  proportion  of   all  unconsented  land  is 


686 


permanently  lost  to  development.  The  manager  has  a  range  of  hypotheses  regarding 
population  response  to  habitat  management.  The  population  is  monitored  during  the 
management  period,  both  to  guide  management  decisions  and  to  reduce  uncertainties 
about  population  dynamics.  In  the  development  below  we  freely  borrow  from 
Nichols  et.  al.  (1995),  whom  we  thank  for  their  lucid  description  of  the  problem. 

Management  options 

In  this  simplistic,  non-spatial  problem,  the  manager  must  decide  how  much  of  the 
available  wildlife  forest  and  open  habitat  to  conserve  for  the  next  five  year  period. 
The  action  chosen  in  period  t  is  denoted  at.  A  policy  A  specifies  a  sequence  of 
management  actions  {a0,  a,,  an}  as  a  function  of  the  observed  population  and 
habitat  conditions. 

Models 

The  models  specify  the  dynamics  of  the  state  variables  through  time.  The  set  of 
population  models  specifies  a  range  of  alternative  hypotheses  about  the  population 
responses  to  management  actions.  The  population  dynamics  and  environmental 
dynamics  are  combined  in  a  single  state  dynamics  equation, 

x  =x  +F(x,aj) 

where  x,  is  a  vector  of  state  variables  including  both  population  status  and  habitat 
status,  z,  is  a  white  noise  process,  and  F-jtxpapZ)  is  the  net  change  in  the  state 
variables  under  model  i. 

We  considered  three  models  describing  population  response  to  management 
actions.  These  models  differ  in  the  degree  to  which  population  dynamics  is  related 
to  the  amount,  composition,  and  spatial  arrangement  of  habitat.  Model  1  was  a 
simple  non-dynamic  habitat  relation  model,  N=dH,  where  iV=species  abundance,  d=a 
constant  density,  //=the  amount  of  habitat  (forest  or  open,  depending  on  the  species). 

Model  2  was  an  equilibrium  source-sink  model  (Pulliam  1988).  For  each  species, 

V1 
N=Hd(\+— — ) 


where  ^species  abundance,  //=amount  of  source  habitat,  ^maximum  density  in 
the  source  habitat,  X7=  population  growth  rate  in  the  source  habitat,  and  \2  = 
population  growth  rate  in  the  sink  habitat. 

Model  3  was  a  dynamic  source-sink  model  (Pulliam  1988).  For  each  species, 
dynamics  were  described  by: 


687 


N(l)t+1  =    [  X(l)N(l)t  if  X(l)N(l)t  <  dH 

\  dH  if  X(l)N(l)t  >  dH 


N(2)t+1  =    [0  if  k(l)N(l)t  <  dH 

\  X(2)N(2)t  +  X(l)N(l)t  -  dH  if  X(l)N(l)t  ;>  dH 

where  7V(l)=population  in  source,  A(l)=population  growth  rate  in  source,  d= 
maximum  density  in  source,  H=  amount  of  source  habitat,  iV(2)=population  in  sink, 
1(2)  =  population  growth  rate  in  sink.  Values  of  the  species  parameters  are  given  in 
Table  1. 


Table  1.- 

-Model  parameters  for  population  dynamics. 

Habitat  affinity 

Kj 

X2 

d 

Species  1 

Forest 

1.3 

1.01 

1 

Species  2 

Forest 

0.85 

0.01 

0.7 

Species  3 

Open 

1.3 

0.85 

9 

Species  4 

Open 

1.01 

0.01 

6 

More  complicated,  spatially-explicit  models  can  and  will  be  constructed,  but  for 
the  present  Models  2  and  3  allow  dispersal  from  favorable  to  unfavorable  habitats, 
and  thus  are  'spatially  informed',  whereas  for  Model  1 ,  in  which  movement  between 
habitats  does  not  occur,  it  is  only  the  amount  of  favorable  habitat  that  determines 
abundance  for  each  species. 

Landscape  dynamics  followed  a  simple  state  transition  model.  Unconsented  land 
was  permanently  lost  to  development  at  a  rate  of  0.049  every  5  year  period.  Of  the 
remaining  habitat,  forest  converted  to  open  at  a  5  year  rate  of  0.26,  and  open  to 
forest  at  a  5  year  rate  of  0.95. 

Components  of  uncertainty 

There  are  four  components  of  uncertainty  which  are  important  in  managing 
biological  systems.  Environmental  stochasticity  is  uncontrollable  variation  in  the 
factors  influencing  population  dynamics.  We  modelled  extinction  as  a  stochastic 
process,  with  probability  distributions  conditioned  on  population  size.  Structural 
uncertainty  reflects  imperfect  knowledge  about  the  relationship  between  management 
actions  and  population  status.  We  explored  the  effect  of  structural  uncertainty  by 
using  a  weighted  average  of  the  three  model  predictions  in  the  optimization 
procedure,  and  varying  the  weights  assigned  to  each  model.  Partial  observability 
refers  to  the  imprecision  in  a  monitoring  program.  Such  imprecision  leads  to 
uncertainty  about  the  appropriate  action  to  take,  and  also  can  affect  the  ability  to 
discriminate  among  alternative  models.  We  plan  to  model  the  effect  of  partial 


688 


observability  by  developing  an  optimal  policy  based  on  complete  observability,  and 
then  using  that  policy  in  a  simulation  model  with  random  variation  around  the 
population  state  variables.  Partial  controllability  refers  to  the  inability  to  implement 
management  decisions  without  error.  We  have  not  considered  partial  controllability 
in  this  exercise,  but  plan  to  do  so  in  future  work. 

Monitoring 

Since  decisions  are  partially  based  on  population  status,  a  monitoring  program  is 
needed.  The  monitoring  program  can  also  serve  to  discriminate  among  the 
competing  models.  The  data  recorded  at  time  /  are  denoted  yn  and  are  related  to  the 
population  according  to 

with  the  random  variable  e  independent  of  z.  The  set  of  all  data  collected  up  to  time 
t  is  denoted  Yt. 

Model  weights 

The  likelihood  at  time  /  that  model  /'  is  the  best  model  relating  population  to 
habitat  is  denoted  p,(t).  The  likelihoods  are  expected  to  change  through  time  as  data 
are  accumulated  and  compared  with  model  predictions. 

Objective  function 

Our  objective  function  Fis  a  weighted  sum  of  a  species  conservation  objective 
and  a  resource  use  objective.  In  practice,  the  weights  would  be  determined  using 
multi-attribute  utility  theory  (Keeney  and  Raiffa  1976).  Our  objective  is  to  maximize 

i 

where  V(A  \  Y(>  Pq)  is  the  expected  value  of  the  objective  for  policy  A  given  the  current 
state  of  knowledge,  ^=0.1875  and  W2  =  0.25  are  weights  for  the  two  objectives, 
S(A)  is  the  number  of  species  at  the  end  of  100  years,  and  C  is  the  proportion  of  the 
available  habitat  conserved  over  all  time  steps.  The  weights  were  chosen  so  that  the 
objective  function  falls  between  0  and  1. 

Solution  method 

All  optimizations  and  simulations  were  done  using  Stochastic  Dynamic 
Programming  (Lubow  1993),  which  uses  backward  iteration  to  find  optimal  policies 
for  discrete  Markov  processes. 


689 


Analysis 


Our  analysis  focused  on  the  effect  of  structural  uncertainty  on  the  optimal  policy 
and  the  expected  value  of  the  objective  function.  Seven  combinations  of  weights 
were  used  (Table  2),  reflecting  a  range  of  likelihoods  ascribed  to  the  three  models. 
The  first  three  cases  reflected  structural  certainty.  In  the  next  three  cases  there  is 
greater  confidence  in  one  model  than  the  other  two,  and  in  the  third  there  is  no  prior 
knowledge  about  the  appropriate  model. 

In  the  extreme  case  in  which  the  data  y,  indicate  with  certainty  which  model  is 
correct,  we  obtain  perfect  information  (Lindley  1985)  about  the  system  structure. 
Given  prior  weights  p,  given  to  the  models,  the  expected  value  of  the  objective 
function  with  perfect  information  is: 

3 

max  V(A)p. 

i=l  A 

The  difference  between  this  value  and  the  maximum  value  of  the  objective  function 
calculated  using  the  weighted  (by  the  p^)  average  of  the  three  models  is  the  expected 
value  of  perfect  information  (EVPI)  (Lindley  1985).  The  EVPI  was  calculated  for 
the  four  combinations  of  weights  which  reflected  structural  uncertainty. 

RESULTS 


Table  2.- 

-The  effect  of  varying  model  weights  on  optimal  decisions  and  objective  V 

Pi 

P2 

Ps 

V  EVPI 

forest  area 
conserved 

open  area 
conserved 

1 

0 

0 

0.46 

75 

25 

0 

1 

0 

0.60 

0 

0 

0 

0 

1 

0.63 

0 

0 

0.7 

0.15 

0.15 

0.47  0.037 

50 

25 

0.15 

0.7 

0.15 

0.56  0.024 

0 

5 

0.15 

0.15 

0.7 

0.57  0.030 

0 

0 

0.333 

0.333 

0.334 

0.53  0.033 

25 

5 

In  general,  the  population  predictions  for  Model  2  were  higher  than  those  for 
Model  1 .  This  was  not  surprising,  since  in  both  models  density  was  proportional  to 
the  amount  of  preferred  habitat,  with  larger  constants  of  proportionality  in  Model 
2.  For  all  combinations  of  state  variables,  the  expected  utility  for  Model  2  was  equal 
to  or  greater  than  the  expected  utility  for  Model  1 .  Similarly,  expected  utilities  for 


690 


Model  3  were  generally  equal  to  or  greater  than  those  for  Model  2.  As  habitat  was 
lost.  Model  2  predicted  immediate  population  equilibrium  with  the  new  conditions, 
while  Model  3  predicted  a  time  lag. 

The  optimal  decisions  also  varied  among  the  models.  In  general,  the  optimal 
amount  to  conserve  was  larger  for  the  model  which  predicted  lower  population  size. 

Varying  the  likelihoods  p;  assigned  to  each  model  affected  both  the  optimal 
decision  and  the  value  of  the  objective  function  for  the  optimal  decision  (Table  2). 

DISCUSSION 

The  degree  of  confidence  placed  in  each  model  affected  both  the  optimal 
management  decision  and  the  expected  value  of  the  objective  function.  Basing 
management  on  an  incorrect  model  without  considering  structural  uncertainty  can 
lead  to  suboptimal  decisions. 

The  EVPI  refers  to  the  expected  increase  in  the  objective  function  from  a 
hypothetical  study  designed  to  eliminate  structural  uncertainty.  For  example,  a  gain 
in  objective  function  units  of  0.033  might  translate  into  an  increase  in  the  expected 
number  of  species  conserved  of  0.033/fF/  =  0.176,  or  an  increase  in  the  proportion 
of  land  open  for  development  of  0.033/fT2  =  0.132.  In  a  realistic  management 
scenario,  data  will  not  discriminate  among  the  models  exactly.  In  that  case,  the 
expected  gain  from  gathering  the  data  depends  on  the  model  likelihoods  (Lindley 
1985).  The  expected  value  of  partial  information  can  be  used  in  an  analysis  in  the 
cost-benefit  tradeoff  of  various  levels  of  monitoring  intensity. 

In  this  exercise,  we  chose  the  model  weights  arbitrarily.  An  adaptive  approach  to 
management  would  involve  treating  the  model  weights  as  state  variables  whose 
dynamics  are  based  on  Bayesian  updating.  In  passive  adaptive  management  (Walters 
and  Hilborn  1978),  information  feedback  into  the  decision  making  process  is  not 
considered  in  evaluating  the  objective  function.  Optimization,  management,  and 
monitoring  are  done  iteratively,  with  model  weights  recalculated  after  each  iteration. 
In  active  adaptive  management,  the  optimization  specifically  considers  the  benefits 
of  future  learning.  Active  adaptive  management  presents  considerable  computational 
difficulties,  because  the  information  state  Yt  increases  in  dimension  through  time.  A 
dynamic  programming  approach  incorporating  future  learning  must  rely  on  a  set  of 
summary  statistics  which  evolve  in  a  Markovian  fashion. 

This  model  exercise  represents  a  preliminary  step  in  developing  adaptive 
approaches  to  conservation  planning.  Our  immediate  objectives  include  exploring 
the  effect  of  partial  observ  ability  on  decision-making  in  our  model  landscape,  and 
performing  similar  analyses  using  population  models  with  spatial  structure.  This 
will  require  some  refinements  in  our  solution  methods  due  to  the  dimensionality 
limitations  of  dynamic  programming.  We  anticipate  taking  an  iterative  simulation- 
optimization  (SO)  approach  to  dealing  with  more  complex  systems.  In  SO,  methods 
such  as  dynamic  programming  or  linear  programming  is  used  to  develop  optimal 


691 


policies  for  simplified  or  static  systems.  More  detailed  simulation  models  are  then 
used  to  track  the  system  dynamics  over  time  using  the  optimal  policy.  If  SO  proves 
to  be  an  effective  strategy  for  dynamic  optimization  in  complex  systems,  we  will 
have  the  analytical  tools  available  for  addressing  optimization  that  is  truly  adaptive, 
in  the  sense  that  the  information  state  is  included  in  the  optimization  procedure. 

ACKNOWLEDGMENTS 

We  thank  Ken  Williams,  Jim  Nichols,  Fred  Johnson,  and  Bill  Kendall  for 
invaluable  consultation  and  input. 

REFERENCES 

Burley,  F.W.  1988.  Monitoring  biological  diversity  for  setting  conservation 

priorities,  pp.  227-230  in  Wilson,  E.O.  (ed.)  Biodiversity.  Washington, 

DC:National  Academy  Press. 
Conroy,  M.J.  and  B.R.  Noon.  Mapping  of  species  richness  for  conservation  of 

biological  diversity:  conceptual  and  methodological  issues.  Ecological 

Applications  (in  press). 
Crocker,  J.E.  and  M.J.  Conroy.  1995a.  A  framework  for  optimal  decision  making 

in  the  management  of  landscapes.  80th  Annual.  Meet.,  Ecological  Society  of 

America,  Snowbird,  Utah.  30  July-3  August  1995.  (Abstract). 
Keeney,  R.L.,  and  H.  Raiffa.  1976.  Decisions  with  multiple  objectives.  New 

York:Wiley. 

Lindley,  D.  V.  1985.  Making  decisions.  New  York: Wiley  207  pp.. 

Lubow,  B.  1993.  Stochastic  Dynamic  Programming  (SDP):  User's  Guide.  Fort 

Collins,  CO  Colorado  Coop.  Fish  and  Wildlife  Res.  Unit. 
Nichols,  J.D.,  F.A.  Johnson,  and  B.K.  Williams.  1995.  Managing  North  American 

waterfowl  in  the  face  of  uncertainty.  Annual  Review  of  Ecology  and  Systematics 

26:177-199. 

Pulliam,  H.  R.    1988.    Sources,  sinks,  and  population  regulation.  American 

Naturalist  132:652-661. 
Scott,  J.  M.,  F.  Davis,  B.  Csuti,  R.  Noss,  B.  Butterfield,  C.  Groves,  H.  Anderson,  S. 

Caicco,  F.  D'Erchia,  T.  C.  Jr.  Edwards,  J.  Ulliman,  and  R.  G.  Wright.  1993.  Gap 

analysis:  a  geographic  approach  to  protection  of  biological  diversity.  Wildlife 

Monographs  123:41pp. 
Walters,  C.J.  1986.  Adaptive  management  of  renewable  resources.  New  York: 

MacMillan.  374  pp. 
Walters,  C.J.,  and  R.  Hilborn.  1978.  Ecological  optimization  and  adaptive 

management.  Annual  Review  of  Ecology  and  Systematics  9:157-188. 
Williams,  B.K.  1989.  Review  of  dynamic  optimization  methods  in  renewable 

resource  management.  Natural  Resource  Modeling  3:137-216. 


692 


Sensitivity  Analysis  of  Species 
Richness  Mapping  to  Variations  in 
Forest-Wildlife  Relationships 

Thomas  W.  Kohley1 

Abstract.  The  Gap  Analysis  Program,  administered  by  the  National 
Biological  Service,  was  implemented  to  identify  "gaps"  in  the  conservation 
of  selected  surrogates  of  biodiversity.  Information  on  habitat  affinities 
and  the  known  occurrence  of  terrestrial  vertebrate  species  are  incorporated 
into  a  GIS  to  develop  maps  of  predicted  species  distributions  and  richness. 
With  the  incorporation  of  species-habitat  suitability  rankings  into  the  Gap 
models,  it  is  possible  to  geographically  delineate  areas  of  critical, 
acceptable,  and  marginally  suitable  habitats  for  each  species.  Further, 
species  richness  can  be  mapped  using  a  "liberal"  or  "conservative" 
approach.  The  liberal  approach  taken  by  Gap  Analysis  contends  that  all 
habitats,  whether  high  or  low  value  for  a  species,  are  of  conservation 
value  and  should  be  included  in  the  measurement  of  species  richness.  The 
conservative  approach  argues  that  only  the  critical  habitats  warrant  the 
highest  conservation  priority  and  all  other  habitat-associations  should  be 
omitted  from  estimates  of  species  richness.  This  research  examined  the 
sensitivity  of  species  richness  to  these  two  mapping  approaches  and 
explored  the  potential  effects  on  biological  conservation  planning.  Results 
indicated  that  the  conservative  approach  differs  substantially  from  the 
liberal  approach  in  terms  of  (1)  the  predicted  representation  of  animal 
species  within  existing  reserves  and  (2)  the  identification  of  species  rich 
hotspots.  


'  GIS  Research  Scientist,  Wyoming  Water  Resources  Center,  Laramie,  WY. 


693 


Development  of  a  Survey  Sampling 
Methodology  for  Rare  Species 

Molly  Van  Caster*      David  Bowden,  Ph.D.1 
Jennifer  Hoeting,  Ph.D.* 


Abstract. — Developing  a  plan  to  find  rare  species  is,  by  na- 
ture, a  difficult  task.  Incorporating  habitat  information  can  im- 
prove sampling  plans.  In  our  current  application,  the  available 
habitat  information  is  rather  sparse  which  increases  the  difficulty 
in  developing  a  useful  sampling  plan. 

The  information  comes  from  past  sightings,  if  they  occurred, 
with  or  without  habitat  information,  scientists'  knowledge,  if  avail- 
able, and  information  on  the  study  area  in  the  form  of  GIS  and 
from  survey  plots.  The  key  is  to  link  the  habitat  information  avail- 
able to  possible  sampling  locations. 

The  model  developed  predicts  likelihood  of  presence  at  a  spe- 
cific site.  The  information  from  the  site  is  used,  as  well  as  habitat 
information  from  the  neighboring  sites.  Neighboring  sites  are  those 
which  have  presence/absence  data  along  with  some  habitat  infor- 
mation. Thus,  sites  similar  to  sighting  locations  would  be  deemed 
more  likely  to  support  the  species.  The  model  and  simulation  stud- 
ies will  be  presented. 


*Ph.D.  Candidate,  Department  of  Statistics,  Colorado  State  University,  Fort  Collins, 
CO. 

^Professor,  Department  of  Statistics,  Colorado  State  University  ,  Fort  Collins,  CO. 
*  Assistant  Professor,  Department  of  Statistics,  Colorado  State  University  ,  Fort  Collins, 
CO. 


694 


Methods  to  Analyse  the  Spatial 
Structure  of  Plant  Communities 

Paul  Braun,  Heiko  Balzter  and  Wolfgang  Kohler1 


Abstract.  -  Though  many  methods  for  analysing  the  spatial  structure  of 
plant  communities  are  available,  few  of  them  deal  with  the  problem  of 
clonal  plant  growth  or  spatio-temporal  dynamics.  This  paper  shows  two 
possibilities  for  quantifying  the  spatio-temporal  dynamics  of  a  plant  po- 
pulation and  a  vegetation.  It  turns  out,  that  fine  shifts  in  plant  popula- 
tions or  plant  community  structure  can  be  discovered,  if  spatial  and 
temporal  dynamics  are  analysed  simultaneously.  Still  there  remains  the 
need  for  improved  or  new  methods. 


INTRODUCTION 

Any  knowledge  about  the  spatial  distribution  of  a  plant  population  allows  to 
draw  conclusions  on  predominant  biotic  or  abiotic  factors.  This  in  turn  enables  us 
to  better  predict  the  population  dynamics  of  a  specific  plant  population  or  even  a 
vegetation. 

If  discrete  objects/individual  plants  are  identifiable,  common  methods  of  quanti- 
tative ecology  are  used  to  further  process  the  data  (Greig-Smith  1983,  Elliott 
1971,  Kershaw  1978,  Upton  and  Fingleton  1990,  Vandermeer  1981).  These  me- 
thods can  be  grouped  as  either  aggregation  or  distance  methods.  While  aggrega- 
tion methods  count  individuals  (e.g.  Variance  to  Mean  Ratio,  Morisita  Index), 
distance  methods  (e.g.  Nearest  Neighbour  Index,  L-Plot)  utilize  inter-plant  distan- 
ces to  conclude,  whether  plants  are  randomly,  clustered  or  regularly  dispersed. 
When  plants  no  longer  can  be  assigned  to  one  point,  as  is  the  case  with  clonal 
plants  or  vegetations,  the  above  mentioned  methods  are  less  useful  to  describe 
spatial  dynamics  of  populations. 

In  vegetation  science  different  procedures  are  used  to  quantify  the  proportions 
of  species  in  a  plant  community,  where  discrete  objects  cannot  be  counted,  e.g. 
Braun-Blanquet-,  Quadrat-,  Point-Quadrat-Method.  Depending  on  the  type  of  da- 
ta, various  methods  can  be  applied  to  detect  spatial  structures.  The  Point-Quadrat- 
Method  provides  a  good  base  to  apply  a  Moving  Window  Algorithm  (Balzter  et 
al.  1995). 

^Justus-Liebig-University,  Dep.  of  Biometry  and  Population  Genetics,  Ludwigstr.  27,  D-35390 
Giessen,  Federal  Republic  of  Germany;  email  gh80(q),agr  ar.uni-giessen.de 


695 


THE  SPATIO-TEMPORAL  DYNAMICS  OF  A  THISTLE-POPULATION 

(CIRSIUM  ARVENSE  (L.)  SCOPOLI). 


After  mapping  plant  populations  on  a  recultivated  area  close  to  Giessen  in  1991, 
we  used  these  data  to  examine  the  usefulness  of  various  aggregation  and  distance 
methods  to  quantify  the  spatial  distribution  of  plant  populations  (Braun  and 
Lachnit  1993,  Braun  and  Lachnit  1994).  We  found  the  Nearest  Neighbour  Index 
(R)  to  be  a  suitable  statistic  for  describing  spatial  and  temporal  changes  in  plant 
populations  of  discrete  objects.  R  depends  on  data  about  the  distance  of  each  plant 
to  its  nearest  neighbour.  As  under  field  conditions  this  information  may  not  easily 
be  obtained,  we  mapped  the  whole  plant  population  (C.  arvense)  in  a  2x2  m  plot. 
With  these  coordinates  the  average  distance,  rM,  of  all  individual  plants  to  its 
nearest  neighbour  had  been  calculated  and  normalized  by  the  expected  nearest 
neighbour  distance,  rE.  This  is  expressed  formally  by 


R  =  rMlrE  [1] 

where  R  is  the  ratio  between  the  measured  average  nearest  neighbour  distance  rM 
and  the  expected  nearest  neighbour  distance  rE  (i.e.  random  distribution  of  indivi- 
dual plants  in  the  population).  The  value  of  rM  is  obtained  by 

;=1 

where  rt  is  the  distance  of  plant  i  to  its  nearest  neighbour  and  TV  is  the  population 
size.  The  expected  value  rE  is  defined  as 


rE 


=  l/(2-V«)  with  u  =  N/A  [3]. 


Here  A  stands  for  the  area  of  the  mapped  plot.  R  close  to  1  indicates  random- 
ness, R  <  1  indicates  a  clustered  and  R  >  1  indicates  a  regularly  dispersed  plant 
population.  To  test  the  significance  of  R,  Clark  and  Evans  (1954)  proposed  the 
following  test  statistic: 

CE  =  R/ge,  where  [4] 

a2£  =(4-n)/(4-tf -H-7i)  [5]. 


CE  is  standardnormally  distributed  and  can  be  tested  by  z- values.  For  six  diffe- 
rent sampling  dates  in  1991  we  calculated  R.  In  Figure  1  the  dynamics  of  the  R 
values  for  the  thistle-population  plot  is  shown. 


696 


N 


~i  1  1  1  1  1  1  1  [  1  1  1  1  1  1  1  1  1  1  r 

60       90       120      150      180      210      240  270 


days  after  1.1.1991 


Figure  1.  -  Spatial  (Nearest  Neighbour  Index)  and  temporal  (population  size)  dynamics  of 
a  thistle-population  in  1991.  Sampling  dates  were:  25.3.,  17.4.,  8.5.,  5.6.,  15.7.  and 
19.9.1991. 


The  population  size  (measured  by  AO  stays  steady  except  for  a  temporary  seaso- 
nal increase.  This  population  dynamics  is  quite  usual  for  natural  populations  in 
central  Europe.  Our  data  suggest  a  decrease  in  clustering  of  the  thistle-population 
during  the  year.  It  seems  that  the  thistle-population  spreads,  although  it  does  not 
increase  in  number.  This  could  be  due  to  competition,  because  grouped  thistle 
plants  experienced  stronger  competition  than  single  plants.  The  observed  loss  in 
aggregation  would  not  have  been  detected  by  only  looking  at  the  population  dy- 
namics. 


THE  SPATIO-TEMPORAL  DYNAMICS  OF  A  MEADOW  VEGETATION 

(L  olio  -  Cy  nosuretum). 

In  the  continuous  case,  i.e.  when  plants  are  not  viewed  as  single  points,  aggre- 
gation and  distance  methods  are  not  efficient.  We  therefore  recorded  an  area  by 
the  Point-Quadrat-Method  (PQM)  and  processed  these  raw  data  with  a  Moving 
Window  Algorithm  (MWA). 

Our  investigation  was  conducted  on  a  meadow  in  Giessen.  Except  for  the  centre, 
which  remained  unmown  all  the  time,  the  meadow  was  cut  up  to  ten  times  a  year. 


697 


This  led  to  an  association  of  Lolio-Cynosuretum  on  the  outer  part  of  the  mea- 
dow (Figure  2).  Whereas  in  the  unmown  centre  only  a  class  of  Molinio-Ar- 
rhenatheretea  could  be  observed  (Oberdorfer  1990). 


Figure  2.  -  Investigated  meadow  in  Giessen  (630  m2).  The  whole  area  is  divided  in  120 
subplots,  which  were  sampled  every  three  months.  The  central  part  remained  unmown  (72 
m2).  The  scientific  name  for  thuja  is  Thuja  occidentalis  and  for  beech  Carpinus  betulus. 

As  shown  in  Figure  2  the  whole  area  was  divided  into  120  subplots.  Inside  of 
each  subplot  a  vegetation  sample  was  taken  with  a  point  quadrat  frame  (three 
needles,  0  1  mm,  each  30  cm  apart).  A  vegetation  sample  means,  that  all  contacts 
of  plant  species  with  any  of  the  three  needles  in  one  frame  were  recorded. 

Four  subplots  or  one  window  half  yielded  a  multivariate  vector,  where  each 
variable  contained  the  mean  number  of  contacts  that  one  plant  species  had  with 
any  of  the  needles.  To  compare  two  successive  years,  the  vegetation  data  of  the 
same  window  half,  but  for  two  successive  years  were  used.  These  two  window 
halfs  -  a  window  -  represented  two  multivariate  vectors  for  which  the  Squared 
Euclidian  Distance  (SED)  was  calculated: 

SEDm=i(xM-XB)2  [6]. 

/=1 


698 


Variables  n  and  w  denote  the  number  and  size  of  the  window,  respectively.  The 
total  species  number  is  s.  Index  A  is  the  window  half  of  the  first  year  and  B  the 
window  half  of  the  second  year.  An  SED  value  of  zero  would  mean  that  no  vege- 
tation change  has  occured.  High  SED  values  indicate  large  vegetation  change 
between  the  years.  In  our  example  we  moved  the  window  stepwise,  that  means 
one  subplot,  from  left  to  right  (Fig.  2).  Thus  the  window  stretched  over  two  years 
but  was  moved  in  space.  For  one  row  this  gave  1 1  values.  After  finishing  one  row 
the  procedure  continued  with  the  next  row.  Thus  the  vegetation  change  between 
years  is  described  by  1 1x9  points,  where  each  value  represents  a  window  compa- 
ring two  years.  Figure  3  shows  the  vegetation  change  from  1993  to  1994. 


Figure  3.  -  Plot  of  the  SED's  between  the  vegetation  from  14.5.1993  and  20.6.1994.  In  the 
background  the  orchard  is  located,  on  the  left  hand  stretches  the  lane,  etc.  (see  Fig.  2). 

There  are  fluctuations  in  the  vegetation  composition,  but  there  is  no  obvious 
change  in  vegetation.  To  get  an  idea  on  the  importance  of  peaks,  we  compared  the 
SED  values  of  1993/94  (Fig.  3)  with  the  corresponding  values  for  1994/95.  (Fi- 
gure 4). 


699 


Figure  4.  -  Plot  of  the  SED's  between  the  vegetation  from  20.6.1994  and  16.6.1995.  In 
the  background  the  orchard  is  located,  on  the  left  hand  stretches  the  lane,  etc.  (see  Fig.  2). 

This  comparison  revealed  a  substantial  change  for  the  period  1994/95.  Indeed 
even  this  big  shift  was  not  easy  to  detect  by  a  routine  inspection  out  in  the  field. 
After  we  knew  these  results  we  searched  for  the  changes  on  the  meadow  and 
found  them.  But  this  shift  might  have  easily  been  overlooked,  without  an  MWA- 
analysis.  This  method  has  proven  to  be  a  very  sensitive  way  to  detect  vegetational 
changes. 

We  asume  that  this  phenomenon  was  the  result  of  massive  herbivory  by  a  local 
rabbit  population  which  augmented  considerable  during  the  mild  winters  of 
1993/94  and  1994/95  in  Giessen. 


CONCLUSIONS 

Our  experience  with  methods  describing  spatio-temporal  dynamics  leads  us  to 
the  conclusion  that  there  are  only  a  few  efficient  methods  to  solve  this  task.  As 
Ripley  (1986)  pointed  out  most  of  the  aggregation  methods  and  some  of  the  di- 
stance methods  can  be  misleading  in  the  interpretation  of  spatial  structures.  For  a 


700 


close  inspection  of  a  spatial  structure  Ripley's  L-Plot  is  recommended.  But  the 
information  density  of  this  statistic  made  us  prefer  R,  which  easily  can  be  com- 
bined with  population  dynamics  data  in  one  plot.  Even  small  changes  in  structure 
are  detectable  thus.  Still  there  is  one  problem  unresolved:  the  form  and  position  of 
structures  in  an  investigation  plot  cannot  be  described.  This  information  might  be 
necessary  to  judge  specific  population  dynamics  based  on  clustered  resources. 
Because  of  this,  but  not  only,  we  recommend  to  always  take  a  look  on  the  raw 
data. 

The  combination  of  PQM  and  M WA  gives  good  estimates  of  even  slight  vege- 
tation, population  or  biomass  changes.  But  one  has  to  keep  in  mind,  that  the  SED 
values  are  relative  values.  This  means  we  can  only  compare  the  distance  between 
specific  window  halfs.  In  other  words,  two  SED's  that  are  plotted  aside  in  Figure 
3  or  4  may  derive  from  different  species  compositions,  even  if  they  have  the  same 
value.  Another  lack  exists  for  the  MWA:  for  our  problem  we  did  not  find  a  stan- 
dard statistical  test  of  significance. 

Although  there  are  still  some  methodological  improvements  needed,  the  study 
of  spatio-temporal  dynamics  of  herbaceous  plant  populations  should  be  intensi- 
fied. 


REFERENCES 

Balzter,  H.  ,  Braun,  P.W.  and  Kohler,  W.  1995:  Detection  of  spatial  discontinui- 
ties in  vegetation  data  by  a  moving  window  algorithm.  In:  From  Data  to  Know- 
ledge: Theoretical  and  Practical  Aspects  of  Classification,  Data  Analysis  and 
Knowledge  Organization.  (Eds.:  Gaul  W.  and  Pfeifer  D.).  -  Springer,  Heidel- 
berg: 243-252. 

Braun,  P.  and  Lachnit,  B.  1993:  Characterizing  the  spatial  distribution  categories 
in  plant  populations.  -  Biometrical  Bulletin  10(3):  25. 

Braun,  P.  and  Lachnit,  B.  1994:  Kennzeichnung  der  raumlichen  Verteilung  von 
Pflanzenpopulationen.  -  Zeitschrift  fur  Agrarinformatik  4:  67-71. 

Clark,  P.J.  and  Evans,  F.E.  1954:  Distance  to  nearest  neighbour  as  a  measure  of 
spatial  relationships  in  populations.  -  Ecology  35:  445-453. 

Elliott,  J.M.  1971:  Some  methods  for  the  statistical  analysis  of  samples  of  benthic 
invertebrates.  -  Freshwater  Biological  Association,  Ambleside.  144  p. 

Greig-Smith,  P.  1983:  Quantitative  Plant  Ecology.  -  Blackwell  Scientific  Pu- 
blishers, Oxford. 

Kershaw,  K.A.  1978:  Quantitative  and  Dynamic  Plant  Ecology.  -  Edward  Arnold, 
London. 

Oberdorfer,  E.  1990:  Pflanzensoziologische  Exkursionsflora.  -  Ulmer,  Stuttgart. 


701 


Ripley,  B.D.  1986:  Spatial  point  pattern  analysis  in  ecology.  In:  Developments  in 
Numerical  Ecology.  (Eds.:  Legendre,  P.  and  Legendre,  L.).  -  Springer,  Berlin: 
407-429. 

Upton,  G.J.G.  and  Fingleton  B.  1990:  Spatial  Data  Analysis  by  Example.  -  Wiley, 

Chichester/New  York. 
Vandermeer,  J.  1981:  Elementary  Mathematical  Ecology.  -  Wiley,  New  York. 


BIOGRAPHICAL  SKETCH 

Paul  Braun  is  agricultural  scientist  in  the  Department  of  Biometry  and  Popula- 
tion Genetics  in  Giessen.  He  obtained  a  Dr.  agr.  degree  at  Giessen  University  in 
1988  and  works  on  agroecology  and  environmental  informatics. 

Heiko  Balzter  is  PhD  student  in  the  Department  of  Biometry  and  Population 
Genetics  in  Giessen.  He  holds  a  Dipl.-Ing  agr.  degree  from  Giessen  University 
since  1 994  and  works  on  spatio-temporal  modeling  in  landscape  ecology. 

Wolfgang  Kohler  is  Biometrician  and  Professor  in  the  Department  of  Biometry 
and  Population  Genetics  in  Giessen  since  1980.  His  research  mainly  concerns 
population  genetics,  environmental  modeling,  experimental  design  and  statistical 
evaluation  of  field  trials. 


702 


Modeling  Population  Dynamics  with 

Cellular  Automata 


Heiko  Balzter,  Paul  Braun  and  Wolfgang  Kohler1 


Abstract:  When  predicting  changes  in  plant  populations,  spatial  inter- 
actions must  be  considered.  Cellular  automata  are  models  incorporating 
both  spatial  and  temporal  dynamics.  Vegetation  data  were  sampled  on  a 
meadow  over  three  years  using  the  point-quadrat  method  and  population 
dynamics  of  three  single  species  were  modeled  with  a  cellular  automaton 
using  stochastic  decision  rules.  Links  with  the  theory  of  Markov  chains 
are  briefly  discussed. 


INTRODUCTION 

Ecological  science  is  structured  in  the  three  major  disciplines  autecology, 
demecology  and  synecology.  While  autecology  examines  the  relation  between  a 
single  organism  and  its  environment,  demecology  looks  at  a  whole  population  of 
organisms  and  synecology  investigates  several  populations  of  different  species 
interacting  with  the  environment.  In  this  sense  our  approach  to  modeling  spatio- 
temporal  population  dynamics  is  an  application  of  statistical  methods  to  a  syn- 
ecological  problem.  Because  there  is  still  a  lack  of  a  complete  theory  of 
succession  (Huston  1994),  the  need  for  building  vegetation  models  to  simulate 
changes  and  patterns  becomes  obvious.  Cellular  automata  provide  a  class  of 
spatio-temporal  models  with  a  simple  basic  structure  but  offer  a  nearly  unlimited 
range  of  possibilities.  Their  value  for  modeling  purposes  is  discussed,  looking  at  a 
vegetation  study  performed  at  the  Agricultural  School  of  Giessen  University, 
Germany,  from  1 993  till  today. 

CELLULAR  AUTOMATA 

A  cellular  automaton  model  consists  of  N  elements,  or  cells,  each  capable  of 
several  discrete  values  in  a  defined  state  space.  The  cells  can  be  linked  in  different 
ways.  In  the  simplest  case  they  are  connected  geometrically  according  to  any 
spatial  order,  such  as  in  a  one-  or  two-dimensional  grid,  in  which  the  spatially 
neighboring  cells  determine  the  state  of  each  cell.  But  the  cells  may  be  connected 
randomly,  too,  such  as  in  neural  networks  (Kauffman  1984).  Here,  only  simple, 

1  Address:  Justus-Liebig-University,  Dep.  of  Biometry  and  Population  Genetics,  Ludwigstrasse  27,  D-35390  Giessen, 
Federal  Republic  of  Germany.  Email:  gha9(w/agr ar.uni-giessen.de 


703 


two-dimensional  automata  with  spatially  symmetrical  neighborhood  links  are 
considered.  Neighborhood  definitions  often  used  are  the  Moore-neighborhood 
consisting  of  the  cell  itself  and  the  eight  neighboring  cells  with  at  least  one  point 
in  common,  and  the  von  Neumann-neighborhood,  that  only  contains  the  cell  itself 
and  the  four  cells  with  at  least  one  side  in  common.  Each  cell  can  change  its  state 
during  discrete  time  steps  t,  according  to  specified  decision  rules  ,  which  can  be 
either  deterministic  or  stochastic  (Czaran  &  Bartha  1992). 

For  a  step  towards  a  theory  of  cellular  automata  see  Wolfram  (1984).  Cellular 
automata  have  a  broad  range  of  applications,  for  example  to  model  chemical 
reactions  with  spatial  diffusion,  the  development  of  spiral  galaxies,  for  phase 
transitions,  crystal  growth,  but  also  quite  often  to  model  biological  and  ecological 
systems  (Wolfram  1983,  Molofsky  1994).  A  literature  review  concerning  ecology 
is  given  by  Phipps  (1992).  Looking  exclusively  at  the  modeling  approaches  to 
ecological  systems  one  furthermore  must  emphasize  the  following  publications. 

Flamm  &  Turner  (1994)  modeled  land  cover  with  a  stochastic  cellular  auto- 
maton integrating  several  data  layers  and  thus  increasing  precision  of  the  model 
output.  In  vegetation  science,  Ratz  (1994)  made  forecasts  of  the  development  of 
spatial  structures  of  boreal  forests  underlying  the  influence  of  fire.  The  effect  of 
fire  on  Banksia  populations  in  Australia  was  also  examined  (Marsula  &  Ratz 
1994).  On  a  small  scale,  Winkler  et  al.  (1994)  modeled  the  dynamics  of  a  dry 
grassland  community,  observing  a  plot  sized  one  square  meter.  A  very  important 
study  was  performed  by  Silvertown  et  al.  1992,  who  modeled  population 
dynamics  of  five  grass  species,  paying  special  attention  to  competition.  They 
extended  the  Markovian  property  to  a  cellular  automaton  model  and  obtained 
results  that  a  non-spatial  model  would  not  have  been  able  to  explain.  A  study  of 
the  impacts  of  a  coal-fired  power  plant  to  freshwater  wetland  in  Wisconsin  was 
carried  out  by  Ellison  &  Bedford  (1995)  with  a  quite  sophisticated  model.  Using 
rank  data  the  model  output  agreed  well  with  the  observed  changes  caused  by 
hydrological  impact.  A  major  problem  of  cellular  automata  models  is  to  formulate 
adequate  decision  rules  determining  the  state  transitions  of  the  cells.  A  good 
solution  to  this  is  presented  in  Wiegand  et  al.  (1994),  who  modeled  five  plant 
species  of  a  semi-arid  shrub  ecosystem  in  South  Africa,  translating  biological  and 
ecological  knowledge  into  decision  rules  depending  on  weather  conditions.  Their 
major  problem  was  the  time  scale  of  the  model,  as  approximately  every  ten  years 
the  seemingly  stable  vegetation  abruptly  changed,  following  a  certain  weather 
event.  However,  it  is  possible  to  successively  improve  the  decision  rules  by  the 
"top-down  approach"  as  described  by  Kummer  et  al.  (1994).  Modeling  the  spread 
of  rabies  they  started  with  a  simple,  one-dimensional  basic  model  and  scaled 
down  in  the  investigation  of  the  system  by  expanding  the  model  spatially  to  two 
dimensions  with  time  steps  of  one  year,  and  then  temporally  to  finer  time  steps  of 
two  months.  Their  experiences  were  satisfying,  for  it  was  possible  this  way  to 
gain  heuristic  insights  in  the  dynamics  of  the  rabies-fox-system. 

2  sometimes  called  "transition  rules"  or  simply  "rules" 


704 


In  a  detailed  discussion  of  cellular  automata  and  ecological  theory,  Phipps 
(1992)  comes  to  the  conclusion  that,  as  far  as  application  to  natural  systems  is 
concerned,  probabilistic  decision  rules  usually  have  a  better  analogy  to  the  system 
than  deterministic  ones,  though  sometimes  their  heuristic  value  is  less. 

Kareiva  &  Wennergren  (1995)  pointed  out  that  a  strength  of  this  class  of 
models  is  that  it  is  capable  of  modeling  ecosystems  with  respect  to  spatial  segre- 
gation, which  often  is  a  condition  for  coexistence  of  predator  and  prey 
populations.  Possibly  cellular  automata  can  contribute  ideas  to  solving  the 
problem  of  competition  and  coexistence  of  plant  species  (Grace  1995). 


MATERIAL  AND  METHODS 


>  ,  ,  > 


Figure  1.-  Point-quadrat 
frame  as  described  by  Kreeb 
(1983)  and  used  in  this  study. 


The  meadow  is  located  in  Giessen,  Germany.  It 
is  regularly  mown  about  ten  times  a  year  apart 
from  an  area  in  its  centre.  The  mown  plant 
community  was  classified  as  I  olio - 
cynosuretum  following  from  the  phytosociolo- 
gical  system  of  Braun-Blanquet  (1964),  whereas 
the  unmown  plant  community  had  a  quite 
different  composition.  For  reason  of  simplicity 
this  difference  in  use  was  ignored  in  our  model. 
The  survey  method  used  was  the  point-quadrat 
method.  An  extensive  statistical  study  of  point- 
quadrat  methods  was  performed  by  Goodall 
(1952).  The  frame  used  in  this  study  holds  three 
pins  above  the  vegetation,  and  during  sampling 

each  pin  passes  through  a  guide  channel  down  to  the  ground  and  the  number  of 
contacts  of  the  pins  to  each  plant  species  is  counted  (figure  1).  In  this  manner  a 
grid  of  10  parallel  transects,  each  consisting  of  12  frame  positions  was  sampled, 
resulting  in  120  subplots.  Methodological  issues  are  discussed  elsewhere  (Balzter 
et  al.  1995).  Data  sets  for  Lolium  perenne  (perenniel  ryegrass),  Trifolium  repens 
(white  clover)  and  Glechoma  hederacea  (ground  ivy)  from  May  1993,  June  1994 
and  June  1 995  were  used. 

The  120  subplots  on  the  meadow  are  used  as  the  cells  of  the  two-dimensional 
cellular  automaton.  When  looking  at  a  certain  plant  species,  it  can  either  be  absent 
(=0)  or  present  (=1)  in  each  cell,  resulting  in  a  dualistic,  discrete  state  space.  First, 
the  transitions  from  1993  to  1994  and  from  1994  to  1995  were  pooled.  In  order  to 
set  adequate  decision  rules,  two  different  approaches  to  stochastic  probability 
estimation  were  then  carried  out: 

1)  Under  the  assumption  that  only  the  state  of  the  cell  itself  at  the  preceding 
time  step  determines  its  actual  state,  a  matrix  of  transition  probabilities  (transition 
matrix)  as  known  from  the  theory  of  Markov  chains  was  estimated  from  the  data. 


705 


Thus,  the  neighborhood"  incorporates  only  the  subplot  itself.  We  will  refer  to 
this  approach  as  „spatial  Markov  chain". 

2)  Assuming  an  influence  of  the  spatially  neighboring  subplots  on  the  subplot 
itself,  the  Moore-neighborhood  (the  cell  itself  and  the  eight  neighboring  cells)  was 
used  in  the  second  set  of  decision  rules.  Transitions  from  any  possible  neighbor- 
hood into  one  of  the  two  states  must  be  considered.  Because  this  would  result  in 
36  different  neighborhood  states,  they  were  classified,  so  that  the  transition  matrix 
is  stochastic  .  Obviously  this  classification  has  to  be  changed  for  frequent  and  rare 
species  (table  1).  In  the  following  text  this  approach  is  called  „Moore- 
neighborhood". 


Table  1.-  Neighborhood  classifications.  Numbers  mean  neighboring  cells  with  species 
present.  Percentage  cover  was  smallest  for  Glechoma  hederacea  ,  followed  by  Lolium  perenne 
and  finally  Trifolium  repens. 


neighborhood  class 

Lolium  perenne 

Trifolium  repens 

Glechoma 
hederacea 

0 

0 

0 

0 

1 

1-2 

1-2 

1 

2 

3-4 

3-4 

2 

3 

5-7 

5-7 

3 

4 

8-9 

8-9 

4-9 

Once  the  transition  matrix  was  estimated,  the  distributions  of  cells  with 
absence  or  presence  of  the  species  at  several  time  steps  were  predicted.  During 
each  time  step  every  cell  randomly  took  a  new  state  depending  on  the 
corresponding  probability  distribution  of  the  transition  matrix. 

RESULTS 

The  transition  matrices  for  the  spatial  Markov  chain  estimated  from  the 
vegetation  data  are  given  in  table  2. 


Table  2.-  Transition  matrices  for  the  spatial  Markov  chain  and  three  plant  species. 


Lolium  perenne 

Trifolium  repens 

Glechoma  hederacea 

from 

to         0  1 

0  1 

0  1 

0  (absent) 

(055  0.45^ 

'0.53  0.47^ 

'0.92  0.08^ 

1  (present) 

U-45  055; 

,0.23  0.77y 

v0.62  0.38; 

For  the  Moore-neighborhood  the  matrices  (table  3)  must  not  be  confused  with 
the  transition  matrices  of  Markov  processes,  because  the  neighborhood  state  space 
is  not  the  same  as  the  cell  state  space,  expressed  in  the  different  numbers  of  rows 


3  A  matrix  is  called  „  stochastic",  ifi)  all  row  sums  equal  one,  ii)  no  element  is  less  than  zero  or  greater  than  one  and  Hi) 
at  least  one  element  in  each  column  differs  from  zero. 


706 


and  columns  of  the  matrices.  Because  of  this  property  is  it  far  from  easy  to 
interpret  the  matrices  in  the  way  we  do  with  the  spatial  Markov  chain. 


Table  3.-  Transition  matrices  for  the  Moore-neighborhood  using  five  neighborhood 
classes  for  the  three  plant  species. 


I,  o limn  pert' ime 

Trifolium  repens 

Glechoma  hederacea 

from  neighborhood 

class 

to          0  1 

0  1 

0  1 

0 

'0.91 

0.09^ 

'1.00 

0.00^ 

'0.99 

0.0  r 

1 

032 

0.48 

034 

0.46 

0.73 

027 

2 

034 

0.46 

030 

030 

0.86 

0.14 

3 

0.46 

034 

027 

0.73 

029 

0.71 

4 

^0.41 

039; 

,0.18 

0.82> 

,0.33 

0.67, 

Figure  2  shows  the  dynamics  of  Lolium  perenne.  The  relative  frequency  of 
cells  with  the  species  present  quickly  reaches  a  limiting  distribution  and  the 
following  variation  is  caused  by  the  limitation  of  the  number  of  cells.  If  an  infinite 
number  of  objects  were  to  pass  the  Markov  chain,  the  limiting  distribution  would 
be  stable.  Interested  in  the  value  of  the  limiting  distribution  we  can  calculate  a 
first  estimate  from  the  last  20  time  steps  /  =  8 1,82,..., 1 00 .  For  Lolium  perenne 
this  results  in  xLp  =  0.50 . 

For  Trifolium  repens  the  predicted 
frequencies  show  similar  results,  but 
stabilize  on  a  higher  level  ( xTr  =  0.67 , 
figure  3)  and  Glechoma  hederacea  tends 
against  xGh  =  0.12  (figure  4). 

The  theory  of  Markov  chains  allows 
us  to  calculate  the  limiting  distributions, 
assuming  that  every  subplot  is  changing 
its  state  according  to  the  Markovian 
transition  matrix.  Because  all  three 
transition  matrices  P  of  the  spatial 
Markov  chains  are  aperiodic  classes  of 
positively  recurrent  states  and  the  state 
space  is  finite,  the  limiting  distributions 
p(co)  are  ergodic  and  can  be  determined 
by  eq.  1  (Heller  et  al.  1978).  Note,  that  the  ergodic  property  means,  that  the 
limiting  distribution  is  independent  of  the  initial  distribution.  In  other  words,  the 
plant  population  reaches  the  same  equilibrium,  wherever  it  starts.  The  results 
(table  4)  agree  to  the  rough  estimates  from  the  data. 

P(od)=  Urn  Pl  [1] 


Figure  2.-  Relative  frequency  of  cells 
colonized  by  Lolium  perenne  over  100  time 
steps  (mean  of  10  simulation  runs) 
predicted  by  the  spatial  Markov  chain. 
x:time;  y:  relative  frequency. 


707 


Figure  3.-  Relative  frequency  of  cells 
colonized  by  Trifolium  repens  over  100  time 
steps  (mean  of  10  simulation  runs) 
predicted  by  the  spatial  Markov  chain. 
x:time;  y:  relative  frequency. 


Figure  4.-  Relative  frequency  of  cells 
colonized  by  Glechoma  hederacea  over  100 
time  steps  (mean  of  10  simulation  runs) 
predicted  by  the  spatial  Markov  chain. 
x:time;  y:  relative  frequency. 


Table  4.-  Ergodic  distributions  of  the  spatial  Markov  processes 


Lolium  perenne 

Trifolium  repens 

Glechoma  hederacea 

state: 
rel.  freq.: 

0  (absent)      1  (present) 
0.50  0.50 

0  (absent)      1  (present) 
0.33  0.67 

0  (absent)      1  (present) 
0.89  0.11 

Using  the  second  set  of  decision  rules,  under  the  assumption  of  the  Moore- 
neighborhood  determining  the  future  state  of  a  cell,  the  model  still  shows 
asymptotical  behavior.  Figure  5(a)  shows  the  results  for  Lolium  perenne. 
Surprisingly  they  do  not  differ  from  those  of  the  spatial  Markov  chain,  as  the 
estimated  limiting  distribution  xLp  =  0.50  is  the  same.  To  examine  the  model 
behavior  two  further  approaches  were  made.  First  the  transition  matrix  was 
applied  to  various  initial  distributions.  All  initial  distributions  tended  to  vary 
randomly  around  the  same  mean  after  a  specific  number  of  time  steps.  Secondly, 
the  cell  grid  of  the  cellular  automaton  was  expanded  to  100  •  100  =  10000  cells. 


(a)  (b) 
Figure  5.-  Relative  frequency  of  cells  colonized  by  Lolium  perenne  over  100  time  steps  as 
predicted  by  the  cellular  automaton  using  the  Moore-neighborhood.  x:time;  y:  relative 
frequency,  (a)  mean  of  10  simulation  runs;  (b)  one  simulation  run,  100x100  cells. 


708 


As  expected  the  variation  was  decreased  strongly  by  this  expansion  (figure 
5(b)),  while  the  mean  of  the  limiting  distribution  is  shown  more  clearly  than  using 
120  cells. 

Looking  at  Trifolium  repens  the  results  are  similar  (figure  6),  but  setting  the 
complete  initial  distribution  to  zero  no  plant  can  ever  establish  itself.  The  reason 
for  this  is  the  transition  matrix  (table  3),  where  a  neighborhood  of  all  zero 
produces  a  cell  in  state  zero  with  probability  pQQ  =  1 .  The  limiting  distribution  of 
Trifolium  repens  is  thus  dependent  on  the  initial  distribution.  Neglecting  this 
special  case,  because  for  practical  purposes  it  is  very  unlikely  that  no  single  plant 
of  Trifolium  repens  would  be  found  on  the  meadow,  the  estimated  limiting  distri- 
bution is  xTr  =  0.72  differing  slightly  from  the  forecast  of  the  spatial  Markov 
chain. 


(a)  (b) 
Figure  6.-  Relative  frequency  of  cells  colonized  by  Trifolium  repens  over  100  time  steps 
predicted  by  the  cellular  automaton  using  the  Moore-neighborhood.  x:time;  y:  relative 
frequency,  (a)  mean  of  10  simulation  runs;  (b)  one  simulation  run,  100x100  cells 


A  completely  different  result  is  obtained  for  Glechoma  hederacea  in  figure  7. 
Although  the  model  behavior  is  still  the  same  as  for  the  other  species,  the  value  of 
the  limiting  distribution  xGh  =  0.66  is  much  greater  than  that  given  by  the  spatial 
Markov  chain.  Which  model  comes  closer  to  reality  can  hardly  be  judged,  but  will 
be  subject  to  further  validation  in  the  following  years.  Because  Glechoma 
hederacea  does  not  typically  cover  two  thirds  of  an  area  the  spatial  Markov  chain 
is  expected  to  be  more  reliable,  but  this  is  only  supposition.  The  importance  of 
selecting  appropriate  decision  rules  based  on  the  right  neighborhood  definition 
can  be  seen  in  this  example.  If  the  neighboring  subplots  do  not  have  a  major 
impact  on  the  species  in  the  subplot,  the  Moore-neighborhood  will  be  the  wrong 
modeling  approach  and  a  spatial  Markov  chain  will  be  preferred.  But,  if  the 
species  is  able  to  colonize  adjacent  subplots  by  stolons  or  seed  dispersal,  the 
Moore-neighborhood  is  likely  to  be  preferable.  Using  stochastic  decision  rules 
whose  probabilities  are  estimated  from  the  data  has  the  major  advantage  that  the 
system  behavior  can  be  modeled  without  knowing  exactly  quantitative  relation- 
ships between  certain  factors  (which  also  would  result  in  a  quite  complicated 


709 


deterministic  model).  The  effects  of  all  major  factors  influencing  the  development 
of  the  population  add  up  to  a  stochastic  probability  density  function  that  can  easily 
be  determined. 


1    17   33  49   65   81  97 


1     17    33   49   65   81  97 


(a)  (b) 
Figure  7.-  Relative  frequency  of  cells  colonized  by  Glechoma  hederacea  over  100  time 
steps  predicted  by  the  cellular  automaton  using  the  Moore-neighborhood.  x:time;  y:  relative 
frequency,  (a)  mean  of  10  simulation  runs;  (b)  one  simulation  run,  100x100  cells 


DISCUSSION 

The  theory  of  cellular  automata  has  not  yet  been  satisfactorily  investigated. 
Despite  the  underlying  simplicity  of  this  class  of  models,  cellular  automata  exhibit 
numerous  different  behaviors.  The  most  important  work  was  done  by  Wolfram 
(1984),  but  there  is  still  a  need  for  mathematical  examinations.  In  our  study  the 
model  behavior  can  be  explained  partly  using  the  theory  of  Markov  chains.  In 
addition,  the  proved  existence  of  a  limiting  distribution  is  important  for  the 
assumptions  of  vegetation  science.  A  limiting  distribution  is  equivalent  to  the 
concept  of  a  climax  of  succession  as  proposed  by  Clements  (1916).  Whether  there 
is  a  climax  of  the  observed  vegetation  changes  or  not  shall  not  be  discussed  in 
detail,  because  various  conditions  necessary  for  the  maintenance  of  the  successio- 
nal  path  are  likely  to  change  before  this  question  arises,  e.g.  the  size  of  the  rabbit 
population.  These  changing  conditions  erect  limits  for  the  predictions  of  the  cellu- 
lar automaton  model.  100  time  steps  are  surely  too  many  to  interpret  in  a  serious 
way,  this  number  was  simply  selected  to  examine  the  long-term  model  behavior. 
But  the  model  output  described  above  probably  gives  a  rough  impression  of  the 
expected  development  of  cover  and  random  variation  of  the  plant  species. 

The  two  cellular  automata  presented  here  show  differences  in  model  output  for 
two  of  the  three  species.  Because  the  only  thing  that  was  changed  was  the 
definition  of  the  stochastic  decision  rules,  the  differences  in  model  output  must  be 
due  to  assumed  spatial  interactions  between  adjacent  subplots.  The  way  in  which 
cells  are  connected  thus  plays  an  important  role.  Furthermore,  the  decision  rules 
used  here  are  very  simple:  They  cover  only  one  single  species  and  do  not  contain 


710 


information  about  biological  conditions  of  the  species,  such  as  mean  length  of 
stolons  or  distance  of  seed  dispersal,  which  were  included  in  the  studies  of 
Wiegand  et  al.  (1994)  and  Ellison  &  Bedford  (1995).  In  fact,  it  is  intended  to 
improve  successively  the  decision  rules  of  our  model  following  the  top-down 
approach  of  Kummer  et  al.  (1994)  by  using  vegetation  data  of  four  time  steps  per 
year,  information  about  the  seed  bank  in  the  soil,  wind  dispersal  and  interactions 
between  species  (the  latter  was  done  for  two  species  by  Balzter  et  al.  1996). 

However,  a  practical  advantage  of  strongly  simplified  decision  rules  in  the 
model  is  that  data  sampling  is  not  as  laborious  as  if  a  lot  of  factors  had  to  be 
measured.  This  feature  is  even  more  important,  if  the  difficulty  of  sampling  spatial 
data  is  considered.  This  may  be  one  reason,  why  over  a  long  time  only  a  few 
scientists  have  applied  spatial  data  analyses  to  their  problems,  although  in  the 
meantime  the  importance  of  both  spatial  and  temporal  structures  and  dynamics  are 
generally  recognized.  In  the  future  more  spatial  data  will  be  usable,  because 
Geographic  Information  Systems  (GIS)  and  large  data  bases  are  becoming  more 
and  more  established.  Linking  cellular  automata  with  raster-based  GIS  seems  to 
be  a  promising  approach.  The  great  freedom  in  setting  up  decision  rules  allows  the 
adaptation  of  a  wide  range  of  knowledge  to  the  model  and  prediction  of  the 
expected  development  under  the  condition  that  the  assumptions  made  in  the 
model  are  true.  This  variety  of  possibilities  might  promote  the  use  of  cellular 
automata  in  future  ecology. 

REFERENCES 

Balzter,  H.  ,  Braun,  P.  &  Kohler,  W.  1995.  Detection  of  Spatial  Discontinuities  in 
Vegetation  Data  by  a  Moving  Window  Algorithm.  In:  Gaul,  W.  &  Pfeifer,  D. 
(Eds.):  From  Data  to  Knowledge:  Theoretical  and  Practical  Aspects  of 
Classification,  Data  Analysis  and  Knowledge  Organization.  Springer,  Berlin, 
243-252. 

Balzter,  H.,  Braun,  P.W.  &  Kohler,  W.   1996  (in  press).  Ein  zellulares 

Automatenmodell  zur  Simulation  der  Vegetationsdynamik  einer  Weidelgras- 

Weifiklee-Weide.  Tagungsberichte  der  AG  Okologie  der  Biometrischen 

Gesellschaft,  Freiburg  &  Leipzig. 
Braun-Blanquet,  J.  1964.  Pflanzensoziologie.  Springer,  Wien  &  New  York. 
Clements,  F.E.   1916.  Plant  Succession.  Carnegie  Institute  of  Washington 

Publication  242,1-512. 
Czaran,  T.  &  Bartha,  S.  1992.  Spatiotemporal  Dynamic  Models  of  Plant 

Populations  and  Communities.  Trends  in  Ecology  and  Evolution  7:2,  38-42. 
Ellison,  A.M.  &  Bedford,  B.L.  1995.  Response  of  a  Vascular  Plant  Community  to 

Disturbance:  A  Simulation  Study.  Ecological  Applications  5:1,  109-123. 
Flamm,  R.O.  &  Turner,  M.G.  1994.  Alternative  Model  Formulations  for  a 

Stochastic  Simulation  of  Landscape  Change.  Landscape  Ecology  9:1,  37-46. 


711 


Goodall,  D.W.  1952.  Some  Considerations  in  the  Use  of  Point  Quadrats  for  the 

Analysis  of  Vegetation.  Australian  Journal  of  Scientific  Research,  B  5,  1-41. 
Grace,  J.B.  1995.  In  Search  of  the  Holy  Grail:  Explanations  for  the  Coexistence  of 

Plant  Species.  Trends  in  Ecology  and  Evolution  10:7,  263-264. 
Heller,  W.D.,  Lindenberg,  H.,  Nuske,  M.  &  Schriever,  K.-H.  1978.  Stochastische 

Systeme.  De  Gruyter,  Berlin. 
Huston,  M.A.  1994.  Biological  Diversity.  The  Coexistence  of  Species  in 

Changing  Landscapes.  University  Press,  Cambridge. 
Kareiva,  P.  &  Wennergren,  U.   1995.  Connecting  Landscape  Patterns  to 

Ecosystem  and  Population  Processes.  Nature  373,  299-302. 
Kauffman,  S.A.  1984.  Emergent  Properties  in  Random  Complex  Automata.  In: 

Farmer,  D.,  Toffoli,  T.  &  Wolfram,  S.  (Eds.):  Cellular  Automata.  Proceedings 

of  an  Interdisciplinary  Workshop,  Los  Alamos,  New  Mexiko.  North-Holland 

Physics  Publishing,  Amsterdam,  145-156. 
Kummer,  G.,  Jeltsch,  F.,  Brandl,  R.  &  Grimm,  V.  1994.  Kopplung  von  Prozessen 

auf  lokaler  und  regionaler  Skala  bei  der  Tollwutausbreitung:  Ergebnisse  eines 

neuen  Modellansatzes.  Verhandlungen  der  Gesellschaft  fur  Okologie  23,  355- 

364. 

Kreeb,  K.H.  1983.  Vegetationskunde.  Ulmer,  Stuttgart. 

Marsula,  R.  &  Ratz,  A.  1994.  Einflufl  von  Feuer  auf  die  Populationsdynamik  von 
serotinen  Pflanzen  -  ein  Modell.  Verhandlungen  der  Gesellschaft  fur  Okologie 
23,  365-372. 

Molofsky,  J.  1994.  Population  Dynamics  and  Pattern  Formation  in  Theoretical 

Populations.  Ecology  75:1,  30-39. 
Phipps,  M.J.  1992.  From  Local  to  Global:  The  Lesson  of  Cellular  Automata.  In: 

DeAngelis,  D.L.   &   Gross.   L.J.   (Eds.):   Individual-based  Models  and 

Approaches  in  Ecology.  Populations,  Communities  and  Ecosystems.  Chapman 

&  Hall,  New  York,  165-187. 
Ratz,  A.  1994.  Modellierung  feuererzeugter  raumlicher  Strukturen  in  borealen 

Waldern.  Verhandlungen  der  Gesellschaft  fur  Okologie  23,  373-381. 
Silvertown,  J.,  Holtier,  S.,  Johnson,  J.  &  Dale,  P.  1992.  Cellular  Automaton 

Models  of  Interspecific  Competetion  for  Space  -  The  Effect  of  Pattern  on 

Process.  Journal  of  Ecology  80,  527-534. 
Wiegand,  T.,  Milton,  S.J.  &  Wissel,  C.  1994.  Ein  raumliches  Simulationsmodell 

fur    eine    Pflanzengemeinschaft    in    der    sudlichen    Karoo,  Siidafrika. 

Verhandlungen  der  Gesellschaft  fur  Okologie  23,  407-416. 
Winkler,  E.,  Klotz,  S.  &  Wissel,  C.  1994.  Mathematische  Modellierung  der 

Phytozonosedynamik  eines  Trockenrasens.  Verhandlungen  der  Gesellschaft  fur 

Okologie  23,451-458. 
Wolfram,  S.  1983.  Statistical  Mechanics  of  Cellular  Automata.  Reviews  of 

Modern  Physics  55:3,  601-644. 
Wolfram,  S.  1984.  Universality  and  Complexity  in  Cellular  Automata.  Physica  D 

10,  1-35. 


712 


Inconsistency  of  Line  Convolution  as 
a  Quality  Factor  in  the  Maintenance 
of  Spatial  Databases 


Michael  J.C.  Weir1 


Abstract.  —  The  interpretation  of  stand  boundaries  on  aerial  photographs 
has  long  been,  and  remains,  an  important  technique  of  acquiring  spatial 
data  to  support  the  management  of  forests  and  forest  land.  As  forest 
management  agencies  move  from  the  initial  establishment  of  spatial 
databases  to  the  task  of  database  updating,  the  maintenance  of  data 
quality  becomes  an  important  issue.  With  time  also,  the  task  of  database 
base  updating  will  pass  to  new  generations  of  photointerpreters.  In 
addition  to  disagreements  in  the  actual  interpretation,  photointerpreters 
differ  in  the  level  of  detail  with  which  they  delineate  stand  boundaries 
and  other  features.  Preliminary  research  has  been  undertaken  to  assess  the 
extent  to  which  photointerpretation  for  database  updating  is  influenced 
by  the  convolution  of  earlier  interpreted  boundaries. 

The  convolution  of  stand  boundaries  independently  delineated  by  three 
experienced  interpreters  on  orthophotographs  of  a  forest  area  in  Austria 
was  measured  and  compared.  Delineations  were  then  made  for  three 
separate  areas,  one  by  each  interpreter.  These  initial  interpretations  of 
orthophotographs  from  1974  were  then  updated  by  the  other  two 
interpreters  using  orthophotographs  from  1984.  Measurement  of  the 
convolution  of  the  initial  and  updated  stand  boundaries  suggests  that 
individual  interpreters  tend  to  retain  a  characteristic  approach  to 
delineation.  Consequently,  it  may  be  expected  that  subsequent  updates  by 
new  generations  of  photointerpreters  may  result  in  databases  containing 
forest  stand  boundaries  delineated  with  inconsistent  levels  of  convolution. 


Lecturer,  International  Institute  for  Aerospace  Survey  and  Earth  Sciences  (ITC), 
Enschede,  The  Netherlands. 


713 


Lead-Based  Paint  Survey  Data: 
Spatial  Variation 

Tze-San  Lee1 

Abstract. — To  identify  and  assess  the  lead-based  paint  (LBP),  Galson 
Corporation  was  retained  by  the  Air  Force  to  conduct  the  LBP  survey  at 
twenty-four  Air  Combat  Command  Bases.  Prior  to  starting  the  survey 
for  each  base,  Galson's  LBP  Survey  Manager,  in  consultation  with  Base 
Civil  Engineering  and  Military  Housing,  reviewed  information  on 
construction  and  painting  history  for  base  housing.  Based  on  this 
review,  housing  types  were  placed  into  different  homogeneous  exposure 
groups  (HEGs).  From  each  HEG,  the  building  and  the  location  of 
painted  components  were  randomly  selected  to  be  tested.  From  twenty- 
four  completed  LBP  surveys,  four  are  randomly  selected  for  an  analysis. 
Our  study  shows  that  due  to  the  inherent  spatial  variation  in  the  LBP 
survey  data  there  is  a  large  base-to-base  variability.  Furthermore,  most 
housing  groups  fail  to  satisfy  the  HEG  assumption.  By  constructing 
choropleth  maps,  homogeneous  subgroups  within  some  nonhomogeneous 
housing  groups  are  then  identified.  Choropleth  maps  for  the  housing 
groups  at  Ellsworth  Air  Force  Base  are  displayed  as  an  illustration. 


'Associate  Professor,  Dept.  of  Mathematics,  Western  Illinois  Univ.,  Macomb,  111.,  61455. 


7U 


Uncertainty  of  Spatially-averag 
Rainfall  Estimates  from  Rain 

Gauges 

Jeffrey  R.  McCollum1  and  Witold  F.  Krajewski2 

Abstract. — This  study  is  conducted  to  aid  the  Global  Precipitation 
Climatology  Project  (GPCP),  which  was  initiated  by  the  World 
Meteorological  Organization.  The  objective  of  the  GPCP  is  the 
production  of  global  rainfall  estimates  by  combining  different 
sources  of  information  such  as  satellite  remote  sensing,  weather 
forecast  models,  and  rain  gauges.  It  is  important  to  understand  the 
accuracy  of  the  rain  gauge  estimates  because  these  are  considered  the 
most  reliable  estimates  for  regions  with  sufficient  numbers  of 
gauges.  Thus,  these  estimates  have  a  significant  influence  on  the 
combined  global  rainfall  products.  Methods  to  model  the  error 
resulting  from  using  the  arithmetic  mean  of  all  the  rain  gauge  values 
in  a  domain  to  estimate  the  spatially-averaged  rainfall  accumulation 
over  the  domain  have  been  developed  previously.  Two  of  these 
methods  are  investigated  in  this  study-  an  empirical  equation  derived 
from  rain  gauge  data  and  an  analytical  equation  derived  from 
statistical  concepts.  Simulations  are  performed  to  determine  the 
accuracy  of  these  two  equations  in  estimating  the  error  of  the  rain 
gauge  mean.  A  stochastic  rainfall  model  is  used  to  simulate  rainfall 
over  large  (256  x  256  krn^)  domains.  


Iowa  Institute  of  Hydraulic  Research,  University  of  Iowa,  Iowa  City,  IA. 
Iowa  Institute  of  Hydraulic  Research,  University  of  Iowa,  Iowa  City,  IA. 


715 


Defining  Urban  Settlements  : 
Certainties  Regarding  Uncertainties 

Mahavir1  and  Sahar  Al-Amir1 


Proper  definition  of  urban  settlements  has  been  a  consistent  problem  with 
city  planners,  administrators  and  census  officers  alike.  The  problem  increases 
manifold,  specially  in  developing  countries,  where  already  complex 
settlements  are  frequently  appended  with  unplanned  and  unexpected 
developments  in  the  fringe  areas.  This  creates  a  'data  poor'  situation. 

Data  obtained  through  aerial  photographs  and  satellite  images  provide  a 
better  understanding  of  this  complexity,  and  at  a  pace  compatible  to  the  rate 
of  change  in  complexity.  Uncertainties  do  remain  in  interpreting  such  data 
depending  upon  the  scale  and/or  resolution.  However,  most  uncertainties  can 
be  traced  back  to  definition  of  such  settlements,  and  a  long  waiting  period 
(say,  10  years  for  a  census)  for  a  settlement  to  be  tested  against  the  definition. 
With  the  conventional  definitions  of  urban  settlements,  uncertainties  will 
remain  even  with  a  100  percent  ground  survey. 

Most  accepted  definitions  of  urban  settlements  are  generally  based  on 
population  size  and/or  function.  There  are  conflicts  in  definitions  adopted  by 
different  authorities  too.  For  instance,  while  the  Census  of  India  bases  the 
classification  on  population,  density  and  employment  in  non-agricultural 
sectors,  the  Governor  of  a  state  may  notify  a  settlement  as  urban  based  on 
generation  of  revenue  also.  Delhi  (in  India)  is  both  a  state  (province)  and  a 
city.  Where  does  one  draw  the  line?  Similarly,  Damascus  (in  Syria)  is  both 
a  province  and  a  city.  Should  the  unplanned  settlements  appended  to  the  main 
city  be  subjected  to  the  development  control  (promotion)  policies  of  the  city 
of  Damascus,  or  the  province  of  Damascus,  or  both,  or  none? 

The  technology  of  remote  sensing  offers  a  tool  enabling  administrators  in 
providing  a  better  understanding  of  this  complexity  in  a  way  that  can  be 
neutrally  executed  by  technicians.  A  flexible  approach  to  definitions,  i.e.,  the 
concept  of  'Continuously  Built-up  Area',  within  a  set  of  rules  for  operation, 
together  with  appropriate  resolution  and  interpretation  techniques  help  to 
minimize  such  uncertainties.  While  low  resolution  data  (e.g.,  72.5m  of  IRS- 
1B)  help  delineating,  the  high  resolution  data  (e.g.,  10m  of  SPOT,  6m  of  IRS- 
1C,  and  aerial  photographs)  support  in  the  operationalization  of  this  concept. 


Division  of  Urban  Survey  and  Human  Settlement  Analysis,  International  Institute  for  Aerospace  Survey  and 
Earth  Sciences,  Enschede,  The  Netherlands. 


716 


Spatial  Dependence  of  Tree  Biomass  in  an 

Old-Growth  Forest 


Franco  Biondi1  and  Donald  E.  Myers2 


Abstract. — Geostatistical  models  were  fitted  to  aboveground  tree  biomass  and 
decadal  biomass  increment  of  an  old-growth,  naturally-seeded  stand  of  southwestern 
ponderosa  pine  {Pinus ponderosa  Dougl.  ex  Laws.  var.  scopulorum)  whose  density 
increased  over  the  twentieth  century.  Our  objective  was  to  determine  the  effects  of 
crowding  on  spatial  dependence  of  tree  biomass,  especially  compared  to  our 
previous  results  concerning  tree  size  and  basal  area  increment  (Biondi  et  al.  1994). 
Despite  increased  crowding,  tree  biomass  maintained  the  same  type  of  spatial 
dependence  from  1 920  to  1 990.  An  isotropic  Gaussian  variogram  was  the  model  of 
choice  to  represent  spatial  dependence  at  all  times.  Tree  biomass  was  spatially 
autocorrelated  over  distances  no  greater  than  30  m,  a  measure  of  average  patch 
diameter  in  this  forest  ecosystem.  Since  patch  diameter  remained  constant  through 
time,  tree  density  increased  by  increasing  the  number  of  pine  groups,  not  their 
horizontal  dimension.  Spatial  dependence  of  tree  biomass  was  greater  than  that  of 
biomass  increment.  Increased  crowding  over  time  altered  spatial  dependence  of 
biomass  increment:  even  though  the  isotropic  Gaussian  variogram  was  still  the 
model  of  choice,  its  range  doubled  from  1920-30  to  1980-90,  while  the  amount  of 
spatial  dependence  was  halved.  The  latter  finding  indicates  that  decadal-scale 
accumulation  rates  over  the  twentieth  century  became  less  dependent  on  mutual  tree 
position.  These  results  confirmed  our  previous  findings  based  on  tree  size  and  basal 
area  increment.  Simulation  models  and  area  estimates  of  stand  dynamics  in  old- 
growth  forests  may  be  improved  by  including  geostatistical  components  to 
summarize  ecological  spatial  dependence.  

REFERENCES 

Biondi.  F.,  D.E.  Myers  and  C.C.  Avery  1994.  Geostatistically  modeling  stem  size  and 
increment  in  an  old-growth  forest.  Canadian  Journal  of  Forest  Research  24:  1354- 
1368. 


Dendrochronologist.  Scripps  Institution  oj Oceanography,  University  oj  California-San  Diego.  La  Jolla.  CA  92093-0215.  LISA 
Professor.  Department  of  Mathematics.  University  of  Arizona.  Tucson.  AZ  85721.  USA 


717 


Sensibility  and  Uncertainty  Analyses 
of  an  Expert  System  to  Determine 
Stand  Treatments 


Otto  ECKMUELLNERl  and  Martin  MOSER2 


Abstract.-  After  introducing  the  case  study  Schlaegl  including  possible 
stand  treatments  the  sense  and  objective  of  sensibility  analyses  for 
determining  a  practical  working  program  is  explained. 

The  success  of  developed  strategies  depends  on  a  clear  formulated  goal 
and  the  accuracy  of  the  used  input  variables.  For  practical  purposes  the 
local  distribution  of  treatments  is  important. 

Local  sensibility  analyses  can  be  used  either  for  determining  the 
consequences  of  decisions  or  for  a  ranking  in  the  urgencies.  Consequently 
treatment  areas  with  similar  strategies  can  be  selected  and  the  financial 
and  technical  need  for  particular  areas  is  predictible.  Furthermore  these 
sensibility  analyses  provide  information  concerning  the  necessarity  and  the 
level  of  success  about  sceduled  strategies. 

Finally  three  examples  of  possible  decision  criterias  are  presented.  The 
first  example  showes  possible  problems  for  different  kinds  of  fertilisation 
and  the  related  pH-value  as  a  used  variable  for  selecting  stand  treatments. 
Then  that  the  sensibility  of  interactions  between  humus  thickness  and  soil 
preparation  is  not  appropriate  for  general  recommandations  because  of  the 
grat  impact  of  unpredictable  local  influences  to  the  result.  The  final 
example  -  Stand  Density  versus  thinning  treatment  shows  a  stabile 
behavior  and  is  therefor  suitable  for  determining  the  urgency  of  thinning 
treatments. 


and   Biometrician,  Institute  of  Forest  Growth,  University  of Bodenkultur  Vienna  Austria 


718 


Stratified  Two  Stage  Sampling 
(Self-Weighted)  for  Assessment  of 
Village  Forest  Resources  in 
Bangladesh 

S.  S.  Islam 

Abstract. — In  1994,  a  study  of  testing  Stratified  Two  Stage  Sampling 
(Self-Weighted)  was  done  in  village  forest  resources  of  14  rural  thanas 
of  Chittagong  district  in  Bangladesh.  A  target  population  of  900  rural 
households  was  formed  through  a  preliminary  survey.  A  field  sheet  was 
developed.  Data  were  collected  as  per  the  field  sheet.  The  seven 
parameters  such  as  total  homestead  area  (ha),  total  area  (ha)  of  compact 
tree  garden,  total  volume  (m3)  of  trees,  total  volume  (m3)  of  compact 
trees,  total  number  of  bamboos,  volume  (m3)  of  compact  trees  per 
hectare  and  number  of  households  with  bamboo  plantation  were 
calculated  through  complete  enumeration  of  the  target  population.  They 
were  also  estimated  on  the  basis  of  250  sample  households  through  the 
proposed  sampling  design.  The  estimates  were  compared  with  the 
population  values  on  the  basis  of  percentage  standard  error,  percentage 
accuracy,  percentage  bias  and  95%  confidence  interval.  Stratified  Two 
Stage  Sampling  (Self-Weighted)  was  found  to  be  a  suitable  inventory 
technique  and  was  recommended  to  apply  in  the  field  for  assessing 
village  forest  resources  in  Bangladesh.  


719 


Modelling  Animal  Movements 
within  a  G.I.S.  Framework 


Joachim  J.,  Janeau  G.  and  Spitz  F.* 


Abstract.-  Connecting  standardised  animal  radio-locations  with  a  geo- 
referenced  environmental  database  may  point  out  various  problems.  Some 
aspects  of  the  methodology  will  depend  on  specific  scientific  goals,  but  the 
main  features  of  the  methodology  will  concern  most  objectives.  A 
Geographic  Information  System  (GIS)  is  built  up  for  the  study  area,  using 
extant  geographical  data  from,  for  instance,  classified  satellite  imagery, 
topographic,  tourist,  ecological  and  vegetation  maps,  administrative  and 
technical  inventories,  fauna,  flora  and  soil  surveys,  in  fact  any  available  geo- 
referenced  environmental  database.  Animal  locations  and  segments  between 
valid  locations  are  vector  data.  They  define  a  vector  coverage  through  which 
all  other  data  have  to  be  taken  into  account.  Raster  and  vector  data  in  the  GIS, 
as  well  as  data  in  the  animal  vector  coverage  are  adjusted  in  both  scale  and 
resolution.  How  to  choose  the  best  scale  and  resolution  for  some  purpose  ? 
The  mere  identity  of  the  studied  species  may  not  be  sufficient  because  spatial 
needs  are  often  different  within  a  same  species  for  breeding,  foraging, 
seeking  or  settling  on  a  territory  etc....  Adaptation  of  choices  to  special 
purposes  are  shown  in  a  few  examples  for  the  Wild  Boar  Sus  scrofa  and  the 
Mouflon  Ovis  gmelini  in  southern  France.  


*  INRA,  Institut  de  Recherches  sur  les  Grands  Mammiferes,  BP  27,  31326 
CASTANET-  TOLOS AN  Cedex  -  France. 


720 


Evaluation  of  Forest  Condition 
Assessment  Data 


F.Mitterbock  *,  O.Eckmullner  ** 


A  central  part  of  forest  condition  monitoring  in  austria  and  europe  is  the  visual  assessment 
of  the  density  of  tree  crowns.  By  no  means  an  objective  parameter,  it's  comparability  over 
several  years  highly  depends  on  trained,  'calibrated'  staff,  which  is  only  available  for  short 
terms  to  contribute  to  special  research  programs  beside  the  national  forest  condition 
monitoring  program.  In  a  case  study  that  develops  concepts  and  means  for  the  restoration 
of  spruce-forest  ecosystems  in  the  austrian  alps,  distribution  patterns  and  timely  changes 
of  crown  densities  provide  basic  data  concerning  the  'where,  when  and  what'  of  specific 
restoration  and  management  measures.  It  is  crucial  for  the  whole  planning  process  to  be 
familiar  with  the  reliability  of  these  data,  because  slight  shifts  in  crown  density  can  result 
in  big  differences  of  area  when  certain  measures  are  to  be  applied.  On  an  area  of  about 
2.000  hectares,  crown  density  on  850  sample  points  was  assessed  by  two  different  student- 
teams  in  1985  an  1991.  This  area-covering  assessment  was  paralleld  by  an  annual 
monitoring  of  crown  density  on  1 6  sample  plots  by  a  trained  and  calibrated  member  of  the 
national  crown  assessment  team.  With  this  calibration  series  on  hand  it  was  possible  to 
check  the  quality  of  the  two  area-covering  assessments,  sort  out  the  real  changes  in  crown 
density  between  1985  and  1991  and  to  show  how  a  specific  felling  policy,  natural 
mortality,  and  the  subjective  bias  of  the  two  different  assessment-teams  contributes  to 
density  dynamics. 


*  GIS-Specialist.  University  for  Natural  Resources.  Vienna.  Austria  . 

Institute  for  Forest  Entomology.  Forest  Pathology  and  Forest  Protection 
**  Biometrician,  University  for  Natural  Resources.  Vienna.  Austria.  Institute  for  Forest  Growth  Research 


721 


Application  of  Remote  Sensing 
Technology  in  Assessing  Village  Wood 
Resources  in  Bangladesh 

Jamil  Ahmed  Chowdhuryl 

Abstract. — Recent  advances  in  the  quality  and  analytical  technology  of 
satellite  imagery  have  made  it  possible  to  achieve  high  precision  in  forest 
resource  inventories.  Increasing  ground  resolution  in  particulars,  is  of 
special  interest  in  the  study  of  small  forest  areas,  and  in  cutting  down 
the  need  for  costly  ground  measurements.  This  paper  presents  the  state- 
of-the-art  in  using  remote  sensing  technology  in  assessing  wood 
resources  in  Bangladesh.  It  presents  some  results  of  forest  inventory 
using  SPOT  imagery,  and  discusses  the  need  to  develop  the  remote 
sensing  technology  towards  the  capability  to  assess  wood  resources  in 
village  conditions  where  mixed  cropping  is  practised.  Practical 
limitations  are  pointed  out.  


INTRODUCTION 

Resent  developments  and  strategies  in  the  forestry  sector  of  Bangladesh  are 
targeted  self  sufficiency  in  wood  resources  through  social/  community  forestry 
practice.  This  represents  a  partial  shift  from  the  traditional  system  of  maintaining 
large,  spatially  separate  and  identifiable  forests  to  a  "balance"  position  where  the 
bulk  of  the  basic  wood  demand  will  be  met  by  wood  from  village  wood  resources. 
The  task  of  the  resource  assessment  expert  now  has  to  shift  towards  estimating 
wood  resources  in  scattered  villages  and  homesteads. 

Preliminary  Sampling  Design 

Each  district  is  stratified  by  crop  density/crop  combinations  and  within  each 
stratum,  a  double  sampling  procedure  is  applied,  the  roles  of  remote  sensing  data 
in  this  scenario  are  : 

-  to  produce  accurate  maps  of  crop  areas. 

-  to  aid  in  crop  stratification 

-  to  identify  the  time/season  suitable  for  crop  measurement. 

In  order  to  reduce  sampling  costs,  it  is  useful  to  consider  a  two  or  three-stage 
design  whereby  the  areas  which  are  to  be  stratified  are  themselves  secondary  or 
tertiary  sampling  units  respectively. 

CONCLUSION 

Establishment  of  a  system  of  plots  comparable  to  the  permanent  sample  plots  in 
villages  might  yield  good  results  if  used  in  combination  with  remote  sensing 
products. 


Divisional  Officer,  Bangladesh  Forest  Research  Institute,  Chttagong,  Bangladesh 


722 


Quality  Evaluation  Services  on  the 

Internet 


Anders  Ostman 


Abstract.  -  Current  methods  for  quality  evaluation  are  often  suited  for 
data  producers  and  not  for  data  users.  Depending  on  type  of  data  and 
production  methods  used,  data  producers  may  utilize  different  quality 
evaluation  techniques  such  as  product  testing,  prototype  testing,  error 
propagation  studies  or  expert  evaluations.  An  ordinary  data  user  is 
however  usually  producing  results  without  having  access  to  proper 
quality  evaluation  procedures. 

A  research  program  have  been  established  at  the  University  of  Lulea, 
Sweden,  aiming  at  developing  quality  evaluation  procedures  for  casual 
GIS  users.  The  approach  taken  here  is  to  design  and  implement 
distributed  and  harmonized  quality  services  in  various  environments. 
These  services  are  then  used  as  building  blocks  in  the  design  of  quality 
evaluation  procedures  for  casual  users.  In  this  paper,  two  initial  studies 
within  this  program  are  presented. 

Error  propagation  studies  and  Monte  Carlo  simulators  are  important  tools 
within  quality  systems  for  spatial  data.  Due  to  the  lack  of  proper  quality 
specifications  of  existing  data  sets,  ordinary  users  are  seldom  using 
proper  quality  evaluation  procedures.  A  Monte  Carlo  simulator  for 
elevation  data  has  therefore  been  designed  and  made  available  on  the 
internet.  The  communication  with  clients  is  currently  based  on  email  and 
the  transfer  ascii  files. 

The  MC  simulator  is  based  on  the  result  of  a  large  international  DTM  test 
conducted  by  the  ISPRS.  It  is  primarly  aimed  for  photogrammetrically 
sampled  digital  elevation  models.  The  client  sends  an  email  to  the  server, 
where  keywords  such  as  flying  height,  point  density,  size  of  grid,  number 
of  simulations  and  email  return  adress  are  specified.  The  server  then 
generates  and  returns  sets  of  error  grids  to  the  specified  email  adress.  The 
values  in  the  error  grids  are  spatially  correlated  according  to  the  results  of 
the  ISPRS  DTM  test  and  specified  parameters. 

Another  quality  service,  which  currently  is  being  designed,  is  test 
procedures  for  quality  evaluation  of  commercial  DTM  software.  Test 
data  from  the  ISPRS  DTM  test  is  here  made  public  together  with  quality 
evaluation  procedures  of  interpolated  terrain  surfaces. 

Further  information  concerning  this  research  program  and  more  detailed 
information  concerning  the  current  quality  services  is  available  at  the 
WWW-site  http ://w w w . luth . se/depts/sb/f o_git/. 


Professor,  Geographic  Information  Technology,  Dept  of  Environmental  Planning  and  Design,  University  of  Lulea, 
S-971  87Luled,  SWEDEN.  Email:  Anders.Ostman@sb.htth.se 


723 


Assessing  Attainment  of  Standard 
by  Two  Stage  Sampling 

Debashis  Kushary 
Rutgers  University  -  Camden 

Abstract:  Consider  assessing  a  process  whether  it  complies  with 
the  regulation  or  not.  Often  it  is  done  in  two  stages  as  described  in 
EPA  (1986)  guide  SW-846  (pp  nine-14  to  nine-17).  In  the  first  stage, 
sample  mean  of  the  measurements  and  then  the  upper  (lower) 
confidence  bound  are  computed  and  compared  against  L,  the 
allowed  maximum  (minimum).  If  L  is  not  in  between  the  mean  and 
the  confidence  limit,  then  appropriate  decision  is  made  by  looking 
at  which  side  L  falls.  If  L  is  in  between  them,  then  a  second  set  of 
samples  is  taken  and  confidence  limit  is  recomputed  using  all  the 
samples  and  again  compared  against  L.  If  100(l-a)%  confidence  is 
desired  in  the  procedure  then  a  different  value  of  a  should  be  used 
to  maintain  the  over  all  error  level  at  a  because  the  first  sample  is 
biased.  Atwood  and  Bryan  (Environmetrics,  1995)  provided  tables 
and  curves  for  practitioners  for  normally  distributed  data.  This 
paper  deals  with  the  nonparametric  solution  to  this  problem  using 
bootstrap  methodology. 


724 


Site  Positioning  Accuracies  Using 
GPS;  Y  Code  Receivers  and  Real-time 

Differential  Options 

Karl  E.  Brown1 

Abstract. — A  majority  of  the  current  nine  DOI  Bureaus  utilize  the  Department  of 
Defense  created  Global  Positioning  System  in  conducting  their  official  activities. 
Accuracies  span  a  continuum  from  multi-  to  sub-meter  positioning.  As  non-military 
users,  civilian  bureaus  have  been  limited  to  the  use  of  the  Standard  Positioning  Service 
(SPS)  with  the  effects  of  Selective  Availability  limiting  [uncorrected]  horizontal  accuracy 
to  approximately  60-100  meters  for  real-time  point  positioning.  Current  differential 
correction  practices  can  correct  field  points  to  1-3  meters  accuracy,  either  after  the  fact, 
or  in  real-time  over  low  power  radio  links,  or  by  subscription  to  a  correction  service. 
Comparatively,  the  Precise  Positioning  Service  (PPS)  delivers  16  meter  horizontal 
accuracy  anywhere  in  the  world,  with  typical  North  American  positioning  experience  at 
10  meters  or  better.  The  authority  to  utilize  the  PPS  was  granted  to  DOI  by  the  Dept. 
of  Defense  (DoD)  on  Aug.  26,  1994.  Unfortunately,  although  commercial  real-time 
broadcast  services  can  allow  a  user  to  obtain  1-3  meters  positioning  accuracy,  a  current 
significant  lapse  in  service  into  many  wildland  areas  does  not  fully  support  the  needs  of 
the  DOI  Bureaus.  Essentially,  these  services  are  based  on  urban  demand  and  urban 
subscription  income,  with  little  evidence  of  future  planning  to  support  users  on  the 
remote  lands  administered  or  studied  by  DOI.  Since  Natural  Resources  and 
Environmental  study  sites  may  occur  in  either  wildland  or  urban  environments,  both 
autonomous  PPS,  and  SPS  receivers  with  real-time  differential  can  meet  the  positioning 
needs  of  DOI  users  needing  to  map,  inventory,  or  obtain  georeferenced  data  on  these 
sites.  Current  studies  for  both  the  US  Coast  Guard  and  the  Federal  Aviation 
Administration  have  proposed  real-time  GPS  augmentation  systems.  Again,  the  proposed 
systems  are  not  sufficient  to  reach  the  wildland  DOI  users  until  full  augmentation  by  the 
FAA  occurs  in  late  1998.  That  full  system  ASSUMES  approval  of  wide  area  correction 
signal  distribution.  The  DoD  maintains  that  such  a  system  might  violate  National 
Security  issues,  and  currently  does  not  support  the  distribution  of  unencrypted  differential 
correctors  over  a  real-time  system.  The  DoD  has  recognized  a  positioning  need  in  the 
civilian  agencies  for  16  meter  accuracy,  and  has  made  the  PPS  and  the  military  PLGR 
receiver  available  to  Federal  civilian  users.  Other  PPS  capable  receivers  are  available 
in  addition  to  the  PLGR;  however,  they  are  more  expensive.  In  order  to  use  the  PPS, 
an  encryption  key  is  utilized,  and  the  receivers  must  be  keyed  annually.  The  National 
Security  Agency  (NSA)  manages  the  keying  materials.  The  keyed  material  is  classified 
as  Secret  [crypto],  and  a  chain  of  custody  with  appropriate  clearances  is  required. 
Typically,  NSA  wants  to  limit  the  number  of  users  of  keyed  material,  so  creating  a  DOI 
cooperative  site  in  1995  would  allow  the  DOI  to  concentrate  it's  access  for  keyed 
material  to  a  secure  site  while  still  delivering  the  needed  capabilities  to  the  field.  A  full 
range  of  accuracies  are  available  to  DOI  users,  depending  on  varying  costs,  hardware  and 
software  installations,  and  training/security  requirements. 

'Nationwide  GPS  Coordinator,  USDI  National  Biological  Service  Denver,  CO 


725 


Statistical  Models  of  Landscape 
Pattern  and  the  Effects  of  Coarse 
Resolution  of  Satellite  Imagery  on 
Estimation  of  Area 

Christine  A.  Hlavka1 


Abstract. — Analysis  of  classified  satellite  imagery  was  conducted  to 
characterize  errors  in  estimates  of  area  based  on  coarse  resolution 
satellite  imagery  which  are  due  to  distortions  in  sizes  of  small  fragments, 
and  to  explore  the  feasibility  of  correcting  for  these  errors  using  a 
statistical  modeling  approach.  Sizes  of  bodies  of  open  water  on  ERS-1 
SAR  and  fire  scars  on  Landsat  MSS  imagery  were  measured.  Statistical 
analysis  of  the  smaller  scars  and  ponds  as  observed  with  this  imagery  of 
relatively  fine  resolution  demonstrated  that  the  distribution  of  the  sizes 
could  be  modeled  by  either  of  two  types  of  statistical  distributions  -  a 
power  distribution  related  to  fractal  processes  or  a  simple  exponential 
distribution.  Comparison  of  the  distribution  of  small  burn  scars  as 
observed  with  Landsat  to  the  distribution  observed  with  AVHRR  showed 
distortions  due  to  the  coarse  spatial  resolution  of  AVHRR  caused  a  net 
overestimation  of  burn  area.  This  bias  was  primarily  caused  by  detection 
in  2  or  3  AVHRR  pixels  of  burns  whose  true  size  was  on  the  order  of 
an  AVHRR  pixel.  


'NASA/Ames  Research  Center. 


726 


Author  Index 


Achard,  F  366 

Ackeret,  J.R  301 

Adler,  P  267 

Al-Amir.  S  716 

Anderson,  M.C  561 

Antle,  A.N  593 

Avena,  G  623 

Baltzer,  H  695,  703 

Besag.  J  493 

Biondi,  F  717 

Bogardi,  1  49 

Bolstad,  P.V  359 

Borgman,  L.F  155 

Boucneau,  G  25 1 

Bowden,  D.C  694 

Bowser.  G  676 

Brand,  G.J  173 

Braun,  P  695,703 

Brion,  P  617 

Brockhaus,  J  401 

Brown,  A  189 

Brown,  K.E  725 

Burgan,  R.E  409 

Cahoon,  D.R.  Jr  383 

Campbell,  J  508 

Canters,  F  225 

Casey,  K  343 

Catanzaro,  D.G  517 

Chen,  X  545 

Chowdhury,  J.  A  722 

Cieszewski,  C.J  649 

Clerke,  W  508 

Cobum,  T.C  4 

Collins,  J.B  121 

Combs,  R.  Jr  359 

Conroy,  M.J  685 

Cressie,  N.A  1 

Crocker.  J.E  685 

Cutler,  D.R  459 

Czaplewski,  R.L  419,409 

Czaplewski,  R.L  467,508 

Dahab,  M  49 

Daosheng,  D  189 

De  Boissezon,  H  375 

Dees,  M  391 

DeGloria,  S  65 

Doihara,  T  545 

Doyle,  P.G  147 

Drummond,  J  189 

Dzur,  R.S  517 

Eastman.  J.R  527 

Eckmullner,  0  718,  721 

Ediriwickrema,  J  401 

Edwards,  T.C.  Jr  459 

Evans,  B.M  561 

Evans,  D.L  419 

Fahringer,  J  508 


Potolp    I  A 

301 

«">\ 

Richer    P  F 

risner,  r.r  , 

..i.  /Z,  JOJ 

Fit7aerald  R  W 

317 

FfKnipht  F  A 

425 

Fnwlpr  Ci  W 

425 

lroll**frr\    F  1 

^OQ 

1  t  o  m    r    \A  F 

S  1  7 

liprtTiPr    i  t  / 

MQ 

vjouerya,  r  



Goodchild,  M.F.  ... 

.217,  568 

finfu/Jiv  i  A 

30 

Gregoire,  T.G.  . 

667 

Griffith  D  A 

147 

lirrtcc    I  _P 

~)f>l 

(in  i  t\  i  rt  A 

AA 1 

finan   R  T 

MQ 

fTiimrv^rty  \A 

401 

frllttfMT\  P 

1  1  ^ 

Hall  1 

1  00 

Hallberg,  G  

129 

t-I  oc^noi  ipr  I— I 

 DO  / 

Ha7elhoff  L 

601 

Uppminl'    A  W 

1  £1 

Hershev  R  R 

73 

Heuvelink,  G.B.M. 

 243 

Hlavka,  C.A. 

 726 

HnrL*  T 

917 

t  c  1  r\t~\t  F 

601 

VA  r\f*t\T\CT  T 

Hofman,  G  

 251 

I-Ir\]  H  Ok  \L'Q  \J  T? 

1  7"* 

Unu/Qrlh    P  T 

01 

Fluahpc    T  P 

1  1  ^ 

t-Iiiicmcin    I  A 

OA'K 

nunier,  kj.j  

717  781 
...L  \  /,  -o  1 

Hutchinson,  M.F.  .. 

701  ^77 

...Ly  1,3// 

7  1  Q 

Tarriiif*?   C~r  \A 

Jdncdu.  \J  

77fl 

 /ZU 

1  Jim  Jin      1  I 

17^ 

Jiang,  H  

 527 

InQ  t*  V» i  m  T 

77ft 

Jones,  K.B  

 560 

467 

kTanoQc  A 

631 

Vprn    ]  W 

1  ^ 

Khnrram  Q 

401 

Kleinn,  C  

 391 

Koebbe,  J  

107 

Kohl,  M  

 441 

Kohler,  W  

695,  703 

Kohley,  T  

693 

Korhonen,  K.T. 

631 

Krajewski,  W.F  

 715 

Kumar,  P  

 99 

Kuo,  W.L  65 

Kushary.  D  724 

Lai,  J.B  610 

Lane.  L.J  147 

Lee,  T-S  714 

Lewis,  A  317 

Lewis,  A  577 

Limp,  W.F  517 

Liu.Z  129 

Lodin,  M  327 

Lowell,  K  235 

Mackey,  B.G  291 

Mahavir  716 

Malingreau,  J-P  366 

Marshall,  P.L  593 

May,  J. A  4 

McCollum,  J.R  715 

McConville,  D  351 

McCulloch,  C.E  65 

McKenney,  D.W  291 

McRoberts,  R.E  659 

Messmore,  J. A  301 

Milliken,  J.A  541 

Mitterbock,  F  721 

Moffett,  J.L  493 

Mohler,  C  65 

Moisen,  G.G  459 

Monserud,  R.A  667 

Moody,  A  83 

Moser,  M  718 

Murillo,  M.L  281 

Myers,  D.E  23.717 

Myers,  W.L  561 

Nasu,  M  545 

Nychka,  D  137 

0"Neill,  R.V  560 

Ohlen,  D.O  409 

Ostman,  A  723 

Owens,  T  351 

Parysow,  P  639 

Patil,  G.P  569 

Pelkie,  C  65 

Pelkki,  M.H  336 

Phillips,  D.L  649 

Pontius,  J.S  434 

Qian,  S.S  164 

Quintanilha,  J.A  501 

Ranneby,  B  45 1 

Reich,  R.M  467 

Ricotta,  C  623 

Riitters,  K.H  560 

Ritman,  K.T  317 

Robey,  M  217 

Rodrigues,  M  501 

Rogowski,  A.S  57 

Royle,  J.A  137 

Rutherford,  B.M  30,39 


727 


Treitz,P.M  91 

Tsai,  F  65 

Turner,  D.P  649 

Van  Caster,  M  694 

van  Elzakker,  C  189 

van  Hees,  W.W.S  553 

Van  Meirvenne,  M  251 


Voyadgis,  D.E  208 

Wade,  T.G  560 

Wadsworth,  R.  199 

Waller,  L.A.  259 

Weinstein,  D.A.  65 

Weir,  M.J.C  309,713 

Wickham,  J.D  560 


Williams,  M.T  477 

Woldt,  W  49 

Woodcock,  C.E.  .121,  535,  541 

Yarus,  J.M  4 

Zhang,  X.F  181 

Zhu,  Z  409 

Zimmerman,  D  129 


728 


Symposium  Steering  Committee 


H.  Todd  Mowrer,  Chair 
Raymond  L.  Czaplewski,  Co-Chair 
James  L.  Smith,  Past  Chair 
Michael  J.C.  Weir,  International  Coordinator 


Science  Advisory  Board 


Joseph  K.  Berry,  Berry  and  Associates 
Harold  E.  Burkhart,  Virginia  Polytechnic  Institute  and  State  University 
Russell  G.  Congalton,  University  of  New  Hampshire 
Noel  A.C.  Cressie,  Iowa  State  University 
George  Z.  Gertner,  University  of  Illinois 
Richard  O.  Gilbert,  Batelle  Laboratories 
Michael  F.  Goodchild,  University  of  California  -  Santa  Barbara 
Gerard  B.M.  Heuvelink,  University  of  Amsterdam 
Christopher  Kleinn,  Universitat  Freiburg 
Barbara  Koch,  Albert-Ludwigs-Universitat 
Michael  Kohl,  Swiss  Federal  Institute  for  Forest,  Snow,  and  Landscape  Research 

Robert  O.  Kuehl,  University  of  Arizona 
Kim  Lowell,  Universite  Laval 
H.  Gyde  Lund,  USDA  Forest  Service  Inventory  and  Analysis 
Ann  Maclean,  Michigan  Technological  University 
Anthony  Olsen,  US  Environmental  Protection  Agency 
G.P.  Patil,  Pennsylvania  State  University 
Denice  Shaw,  US  Environmental  Protection  Agency 
J.P.  Skovsgaard,  Danish  Forest  and  Landscape  Research  Institute 
Steve  Stehman,  University  of  New  York  -  Syracuse 
Mike  Story,  National  Biological  Service 


The  United  States  Department  of  Agriculture  (USDA)  prohibits  discrimination  in  its 
programs  on  the  basis  of  race,  color,  national  origin,  sex,  religion,  age,  disability, 
political  beliefs  and  marital  or  familial  status.  (Not  all  prohibited  bases  apply  to  all 
programs.)  Persons  with  disabilities  who  require  alternative  means  for  communica- 
tion of  program  information  (braille,  large  print,  audiotape,  etc.)  should  contact  the 
USDA  Office  of  Communications  at  (202)  720-2791. 

To  file  a  complaint,  write  the  Secretary  of  Agriculture,  U.S.  Department  of 
Agriculture,  Washington,  D.C.  20250,  or  call  (202)  720-7327  (voice)  or  (202)  720-1 1 27 
(TDD).  USDA  is  an  equal  employment  opportunity  employer. 


Great 
Plains 


U.S.  Department  of  Agriculture 
Forest  Service 

Rocky  Mountain  Forest  and 
Range  Experiment  Station 


The  Rocky  Mountain  Station  is  one  of  eight 
regional  experiment  stations,  plus  the  Forest 
Products  Laboratory  and  the  Washington  Office 
Staff,  that  make  up  the  Forest  Service  research 
organization. 

RESEARCH  FOCUS 

Research  programs  at  the  Rocky  Mountain 
Station  are  coordinated  with  area  universities  and 
with  other  institutions.  Many  studies  are 
conducted  on  a  cooperative  basis  to  accelerate 
solutions  to  problems  involving  range,  water, 
wildlife  and  fish  habitat,  human  and  community 
development,  timber,  recreation,  protection,  and 
multiresource  evaluation. 

RESEARCH  LOCATIONS 

Research  Work  Units  of  the  Rocky  Mountain 
Station  are  operated  in  cooperation  with 
universities  in  the  following  cities: 


Albuquerque,  New  Mexico 
Flagstaff,  Arizona 
Fort  Collins,  Colorado* 
Laramie,  Wyoming 
Lincoln,  Nebraska 
Rapid  City,  South  Dakota 


'Station  Headquarters;  240  W.  Prospect  Rd.,  Fort  Collins,  CO  80526